Project Thesis

INTRACRANIAL HEMATOMA DETECTION

INTRACRANIAL HEMATOMA DETECTION

Undergraduate graduation project Report submitted in partial fulfillment of the requirements for the Degree of Bachelor of Science of Engineering in Department of Electronic & Telecommunication Engineering University of Moratuwa

Project Group

Supervisor

G.K.I. Abayarathna 050001A G. Gartheeban 050131V E.D.R. Kumara 050234N W.M.D. Soysa 050440R

Dr. A. A. Pasqual

May 19, 2009

ii

Approval of the Department of Electronic and Telecommunication Engineering

Head, Department Of Electronic and Telecommunication Engineering

This is to certify that we have read this project and that in our opinion it is fully adequate, in scope and quality, as an Undergraduate Graduation Project.

Supervisor: Name and Signature

Date:

iii

Abstract THESIS TITLE Supervisor: Dr. A. A. Pasqual Keywords : intracranial hematoma detection, near infra-red, near infra-red spectroscopy, iHD Intracranial hematoma is a treatable potentially fatal secondary injury, with 80 percentage of survival rate if identified and treated timely. Traumatic brain injury, the most prevailing cause of fatality in an accident, is one of the prime causes of intracranial hematoma, along with post-surgery complications. The nature of the causes requires portability, ability to mass produce, affordability in detection methods. However conventional detection technologies such as CT scan and MRI, albeit being accurate and comprehensive, do not offer the aforementioned features. A portable, affordable, safety abiding device to detect intracranial hematoma has rich set of applications such as quick diagnosis, especially on patients with no external wounds, wider availability, by making the devices available in virtually every infirmaries, pre-scanning before enlistment for CT scans, on-site detection, especially on battle fields, nationwide catastrophes, etc. This report describes the design and development of the intracranial hematoma detector (iHD) that detects the presence of the hematoma with above reasonable accuracy. The solution consists of the iHD terminal and iHD mobile application (iHDMA). iHD terminal consists of Near Infra-red (NIR) LED and sensor system to measure the absorption of NIR light. The measured values are encoded and transmitted, through BluetoothT M , to a mobile computer running iHDMA . In iHDMA, using, optical density (OD) calculations, threshold detection, post detection integration and bidirectional associative memory, the presence of hematoma is ascertained. In addition to reporting OD and probability estimation, iHD system can detect the hematomas, semiautomatically for a given constant probability of false alarm. The system is currently accurate enough to detect Extracerebral Hematoma (EH) that could cause and absorbtion difference above 0.2 in optical density; it the same category that needs immediate attention as well.

iv

“To our parents, teachers and friends.”

v

ACKNOWLEDGEMENTS Dr. Ajith Pasqual has been a supportive and encouraging project advisor. He showed a strong commitment and was very enthusiastic in extending a helping hand. He was very approachable and open, on many occasions we were able to contact him even in non-office hours, and notably one time, when he was on a tour. Dr. Pasqual supervised two projects this year, and amidst his involvements in numerous activities, he has always been there for us. In the middle, even when we were about to follow through another project with another supervisor, he was more than helpful. We thank him for being an excellent mentor and a guide. Prof. J.A.K.S. Jayasinghe has been helpful is obtaining the enclosure within short span of time. He, albeit his heavy schedule allocated time to check the design and supervise the operation of the Rapid Prototyping Machine. We thank him for his assistance. Mr. Kithsiri Samarasinghe, former Head of the Department, offered valuable advices as a supervisor, mentor, lecturer, and Head of the Department. His popular ”five minute introduce yourself”, triggered many of us to reflect on ourselves and encouraged us to be forward and assertive. More importantly his advices on final year project procedures, engineering aspects and industry oriented preparation were invaluable. We thank him for being a wonderful teacher. Dr. Chulantha Kulasekare, Head of the Department, has served in the supervisory panel during feasibility study presentation and offered valuable advices. Further, during the project selection period, when we approached him with our idea, he was very supportive and encouraging. He was clairvoyant in recognizing the potential pitfalls and readily warned us about the caveats. We thank him for being a supportive educator. We are in debt to Dr. Kosala Ranatunga for proposing the idea, arranging meetings with the Director of the Accident ward, General Hospital, Colombo and other surgeons, and taking all the trouble in obtaining permissions. He was very supportive during the last phase of the project and visited on the 17th of May to vi

inspect the operation of the device as well. We thank him for his facilitation. We also thank Dr. Himashi Kularathne, neurosurgeon, Head of Neural Surgical Ward, General Hospital, Colombo for granting us permission to conduct field tests on patients undergoing surgery for extracerebral hematoma. Mr. Udaya Chinthaka Jayatilake, lecturer, Department of mathematics, is a proficient mathematician who is keen on exploring the impossibles. He introduced us Artificial Neural Networks and Bidirectional Associative Memory. His lectures were simple and informative, always looking for ways to integrate with practical applications. We thank him for being a guide to elusive mathematics. A number of people have helped us in iHD project. Sashitha Nalin, from Department of Mechanical engineering, was helpful in designing the chassis. Few other batch mates offered their valuable advice throughout the project. We also thank the staff from Engineering Design Center for letting us use Rapid Prototype machine and their help in using it. We also thank other staff members, technical assistants, and non-academic staff for all their support. We also thank our parents and friends for being there, during our tenure at University of Moratuwa, supporting us financially and otherwise, and tolerating us during the hard times, especially during the busy days.

vii

TABLE OF CONTENTS

APPROVAL

iii

ABSTRACT

iv

DEDICATIONS

v

ACKNOWLEDGEMENTS

vi

TABLE OF CONTENTS

viii

LIST OF FIGURES

xii

LIST OF TABLES

xiv

ABBREVIATIONS

xv

1 INTRODUCTION 1.1 Intracranial Hematoma Detector (iHD) . . . . . . . . . . . . . . . . 1.2 Organization of the Report . . . . . . . . . . . . . . . . . . . . . . . 2 LITERATURE SURVEY 2.1 Intracranial Hematoma (IH) . . . . . . . . . . . . 2.1.1 Types of Intracranial Hematoma . . . . . 2.1.2 Traumatic Brain Injury (TBI) . . . . . . . 2.2 Other diagnosis procedures . . . . . . . . . . . . . 2.2.1 Computed Axial Tomography (CAT) . . . 2.2.2 Magnetic Resonance Imaging (MRI) . . . 2.3 Near Infrared Spectroscopy (NIRS) . . . . . . . . 2.3.1 Application of near infrared spectroscopy hematoma detection . . . . . . . . . . . . 2.4 Safety Regulations . . . . . . . . . . . . . . . . . 2.5 Algorithms . . . . . . . . . . . . . . . . . . . . . . 2.5.1 Threshold detection . . . . . . . . . . . . . 2.5.2 Post detection integration . . . . . . . . . 2.5.3 Bidirectional Associative Memory (BAM) 2.6 Chapter Summary . . . . . . . . . . . . . . . . .

viii

. . . . . . . . . . . . . . for . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . intracranial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 3 5

. . . . . . .

6 7 7 9 10 10 10 11

. . . . . . .

12 14 15 15 15 16 16

3 SYSTEM OPERATION 3.1 Using the system . . . . . 3.2 Operation of iHD terminal 3.3 Operation of iHDMA . . . 3.4 Chapter Summary . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

4 SYSTEM ARCHITECTURE 4.1 iHD System Architecture . . . . . . . . . . . . . . . . . . . . . . . 4.2 iHD Hardware Architecture . . . . . . . . . . . . . . . . . . . . . 4.2.1 Signal generation and Sensor Subsystem . . . . . . . . . . 4.2.1.1 Pulse generation . . . . . . . . . . . . . . . . . . 4.2.1.2 NIR light emission . . . . . . . . . . . . . . . . . 4.2.1.3 Signal reception, amplification and Analog to Digital Conversion (ADC) . . . . . . . . . . . . . . . 4.2.1.4 Interfacing and Transmitter - Receiver Separation 4.2.2 Processing & Controlling Subsystem . . . . . . . . . . . . 4.2.2.1 PIC Microcontroller . . . . . . . . . . . . . . . . 4.2.2.2 Sensor Interfacing . . . . . . . . . . . . . . . . . 4.2.2.3 Possible methods of computation . . . . . . . . . 4.2.3 Communication Subsystem . . . . . . . . . . . . . . . . . . 4.2.3.1 UART . . . . . . . . . . . . . . . . . . . . . . . . 4.2.4 Power Management Subsystem . . . . . . . . . . . . . . . 4.2.4.1 Low Dropout (LDO) Regulators . . . . . . . . . 4.2.4.2 Power Source . . . . . . . . . . . . . . . . . . . . 4.2.4.3 DC Power and Charging . . . . . . . . . . . . . . 4.2.4.4 Sleep, Power Saving Modes . . . . . . . . . . . . 4.2.5 Input Output Subsystem . . . . . . . . . . . . . . . . . . . 4.2.6 Safety concerns . . . . . . . . . . . . . . . . . . . . . . . . 4.3 iHD Firmware Architecture . . . . . . . . . . . . . . . . . . . . . 4.3.1 Communication state implementation . . . . . . . . . . . . 4.3.2 Data Acquisition state implementation . . . . . . . . . . . 4.3.3 User Indication IO . . . . . . . . . . . . . . . . . . . . . . 4.4 Communication Protocol . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.3 Message Starting Header . . . . . . . . . . . . . . . . . . . 4.4.4 Message Body . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.5 Message-ending Header . . . . . . . . . . . . . . . . . . . . 4.4.6 Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 HEMATOMA LIKELIHOOD ESTIMATION IN iHDMA ix

. . . .

17 17 21 22 25

. . . . .

26 26 27 28 29 29

. . . . . . . . . . . . . . . . . . . . . . . . . . .

31 33 36 36 37 38 39 40 42 42 43 43 44 44 45 45 46 46 46 47 47 48 48 49 49 49 49 50

5.1 5.2 5.3 5.4 5.5 5.6 5.7

Optical Density calculation on Isolated Scans . 5.1.1 Merits and Demerits . . . . . . . . . . . Characteristics of Intensity values and OD . . . Reference Estimation through normalization . . 5.3.1 Moving average based normalization . . Threshold detection . . . . . . . . . . . . . . . . Post detection integration . . . . . . . . . . . . Application of Bidirectional Associative Memory Summary . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (BAM) . . . . .

6 RESULTS, EVALUATION AND DISCUSSION 6.1 Testing procedures and clinical trials . . . . . . . 6.1.1 Trial Objectives and Purpose . . . . . . . 6.1.2 Trial Design . . . . . . . . . . . . . . . . . 6.1.3 Selection and Withdrawal of Subjects . . . 6.1.4 Treatment of Subjects . . . . . . . . . . . 6.1.5 Assessment of Efficacy . . . . . . . . . . . 6.1.6 Assessment of Safety . . . . . . . . . . . . 6.1.7 Statistics . . . . . . . . . . . . . . . . . . . 6.2 Evaluation and Discussion . . . . . . . . . . . . . 6.2.1 Sensor outputs . . . . . . . . . . . . . . . 6.2.2 Performance metrics . . . . . . . . . . . . 6.3 Presentation of results . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . .

50 51 51 52 52 55 57 58 59

. . . . . . . . . . . .

60 60 60 60 61 61 62 62 62 63 63 63 64

7 CONCLUSION 69 7.1 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 7.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 7.3 Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 BIBLIOGRAPHY

72

APPENDICES

77

A iHD SCHEMATIC

77

B COMMUNICATION PROTOCOL SPECIFICATION B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . B.1.1 Structure . . . . . . . . . . . . . . . . . . . . . . B.2 Message-starting Header . . . . . . . . . . . . . . . . . . B.3 Message Body . . . . . . . . . . . . . . . . . . . . . . . . B.4 Message-ending Header . . . . . . . . . . . . . . . . . . . B.5 Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.6 Example Word Patterns . . . . . . . . . . . . . . . . . . x

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

79 79 80 81 81 82 82 83

B.6.1 Word Types . . . . B.6.2 Message ID . . . . B.6.3 Command or Reply B.6.4 Frame ID . . . . . B.7 Summary . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

83 84 84 84 85

C DATASHEETS

86

D BIDIRECTIONAL ASSOCIATIVE MEMORY (BAM)

91

xi

LIST OF FIGURES 1.1 1.2

Intracranial Hematoma detector . . . . . . . . . . . . . . . . . . . . Banana-shaped light path . . . . . . . . . . . . . . . . . . . . . . .

2.1 2.2 2.3 2.4

Traumatic Brain Injury . . . . NIR Absorption . . . . . . . . Optical Density Histogram . . Significance of Optical Density

. . . . . . . . . . . . . . . . . . variations

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. 9 . 12 . 13 . 14

3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8

Modes of operations of iHD terminal . . Optical Density Scan Positions . . . . . . Visualizing Results . . . . . . . . . . . . Complete Analysis . . . . . . . . . . . . iHDMA capture operation . . . . . . . . iHDMA store and data-process operation iHDMA advanced analysis operation . . iHDMA configurations . . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

18 19 20 20 22 23 24 24

4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 4.11 4.12

iHD System . . . . . . . . . . . . . . . . . . . . . iHD Hardware Architecture . . . . . . . . . . . . Burst of Pulses . . . . . . . . . . . . . . . . . . . OPT101 Sesor . . . . . . . . . . . . . . . . . . . . OPT101 High sensitive light to voltage converter Dark Current Offset Correction . . . . . . . . . . NIR LED and Sensor attachment . . . . . . . . . Flexible Elastic Optical Probe . . . . . . . . . . . Bluetooth Module . . . . . . . . . . . . . . . . . . Bluetooth Integration via UART . . . . . . . . . Bluetooth Stack . . . . . . . . . . . . . . . . . . . Firmware Architecture . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

27 28 29 32 32 33 34 35 39 41 42 45

5.1 5.2 5.3 5.4 5.5

Intensity Samples (Before Normalization) Intensity Samples (After Normalization) Dynamic threshold . . . . . . . . . . . . Results after threshold detection . . . . . Post detection integration . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

53 53 56 56 57

6.1 6.2

Reference samples before normalization . . . . . . . . . . . . . . . . 64 Reference samples with normalization factor 4 . . . . . . . . . . . . 65 xii

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

4 4

6.3 6.4 6.5 6.6 6.7 6.8 6.9

Reference samples with normalization factor 9 . Reference samples with triangular normalization Reference samples with triangular normalization OD calculation on scan and reference intensities BAM to filter out unlikely cases . . . . . . . . . Visualization of the results. . . . . . . . . . . . Presentation of Quick scan OD output . . . . .

. . . . . function function . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

65 66 66 67 67 68 68

A.1 iHD SCHEMATIC 1 . . . . . . . . . . . . . . . . . . . . . . . . . . 77 A.2 iHD SCHEMATIC 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 78 B.1 B.2 B.3 B.4 B.5 B.6 B.7 B.8 B.9 B.10 B.11

Typical Message . . . . . . . . . . . . . . . . . . . . . Message-starting Header . . . . . . . . . . . . . . . . Types of Message Starting Headers (Commands) . . . Types of Message Starting Headers (Replies) . . . . . Message Body of Ping Reply . . . . . . . . . . . . . . Message Body of Update settings and request settings Message Body of Request Data . . . . . . . . . . . . Message-ending Header . . . . . . . . . . . . . . . . . Frame-idle . . . . . . . . . . . . . . . . . . . . . . . . Data-Frame . . . . . . . . . . . . . . . . . . . . . . . Typical word . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . reply . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

80 81 81 81 82 82 82 82 82 83 83

C.1 C.2 C.3 C.4 C.5

Datasheet - PIC 18F452 . . . . . . Pin Diagram - PIC 18F452 . . . . . Specifications of OPT101 . . . . . . OPAMP characteristics of OPT101 Datasheet - NIR LED . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

86 87 88 89 90

xiii

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

LIST OF TABLES 4.1 4.2 4.3

iHD terminal power budget . . . . . . . . . . . . . . . . . . . . . . 42 Message Starting Header . . . . . . . . . . . . . . . . . . . . . . . . 48 Types of message commands . . . . . . . . . . . . . . . . . . . . . . 48

xiv

ABBREVIATIONS

ADC

Analog - to - Digital Converter

ANN

Artificial Neural Network

APD

Avalanche Photodiodes

BAM

Bidirectional Associative Memory

BT

Bluetooth

CAT

Computed Axial Tomography

CCD

Charge Coupled Devices

CT

Computer Tomography

EEPROM Electrically Erasable Programmable Read-Only Memory EH

Extracerebral Hematoma

FDA

Food and Drug Administration

IH

Intracranial Hematoma

iHD

intracranial Hematoma Detector

iHDMA

intracranial Hematoma Detector Mobile Application

LCD

Liquid Crystal Display

LED

Light Emitting Diode

MRI

Magnetic Resonance Imaging

NIR

Near Infra-Red

NIRS

Near Infra-Red Spectroscopy

OT

Optical Topography

PDA

Personal Digital Assistant

PMT

PhotoMultiplier Tubes

PWM

Pulse Width Modulation

SiPD

Silicon Photodiodes

SPP

Serial Port Profile

TBI

Traumatic Brain Injury

xv

Chapter 1 INTRODUCTION

Medical diagnosis is the process of determining the presence of a disease in an individual, followed by the confirmation through analysis and past experience of similar observations [1]. Technology has enabled us to utilize elaborate equipments to diagnose with greater accuracy and precision, in addition to conventional physical examinations, with the assistance of Medical Technologists [2]. Nonetheless our dynamic lifestyles, exorbitant price of medical equipments, exiguity of advanced equipments in developing countries and remote areas, and lack of technical expertise have always been in favor of portable simple on-site detection technologies. Intracranial hematoma detection is a medical diagnosis of determining the presence of hemorrhage inside the cranium, which could potentially lead to death if not detected and treated timely. While advanced high-priced detection methods like Computed Tomography (CT) [3] is widespread today, we are motivated by the promise of affordable portable technologies which could benefit people living in developing countries like Sri Lanka. In addition, a portable device is highly desirable in on-site detection and thus unveil another set of applications. A potential list of such applications are, • Quick diagnosis : Patients with no external wounds are likely to be overlooked and hence might result in inadvertent severity. In addition, patients with inconspicuous and latent symptoms might also suffer the same fate. Further, due to overlapping symptoms it is increasingly becoming common 1

to be misdiagnosed when sophisticated tests are not carried out [4]. In developing countries like Sri Lanka where the medical tests are costly and equipments are scanty, this is not a rare case. Wider availability of an affordable, less cumbersome device to provide positive confirmation of the disease will be of great value to a physician in making a well informed decision (Courtesy - Dr. E. C. Kulasekare) • Wider availability: By making Intracranial Hematoma Detector (iHD) available in virtually every infirmary, a quick diagnosis can be performed and on positive indication they could be referred to teaching hospitals with adequate facilities for additional tests and further treatment. • Pre-scanning before enlisting to CT scans : CT equipments are prohibitively expensive and hardly affordable by a standard hospital in developing countries such as Sri Lanka. This leads to a long waiting list for the usage that results in unacceptable delays and deaths that could have been otherwise prevented. Such a device could be used as a preliminary measure to filter out the patients by confirming the presence of hematoma. • On-site detection : During a public accident or nationwide catastrophe not everyone needs the same medical attention. While people with benign external wounds could wait post first-aid treatment, those who are with urgent medical conditions should be transferred and attended immediately. However in the absence of an external wound, Traumatic Brain Injuries (TBI), a major cause of death and disability worldwide [5], are not detected and hence lead to fatalities. This is common in battle fields, accidents in sports and adventure and industrial mishaps. An on-site detection technology addresses the particular issue at hand. Observing the distinctions between traditional equipments and portable devices, it becomes clear that they are in fact complementing each other rather than substituting, thus improves the outcome together. Due to the fundamental difference 2

between the underlying technology, otherwise-would-have-been-trivial factors such as usability, noise treatment, power consumption, etc become significant. Therefore, unusual measures have to be undertaken. In determining the successfulness of solution, the following are deemed to be critical: • Safety assessment:

Since it is a medical instrument that directly deals

with patients it should abide by the safety regulations and not become the cause for further complications. Primary concerns are heat generation and exposure to radiation. • Reliable: The results should reflect the actual condition with high probability and device must be capable of self learning to adjust its parameter to improve the accuracy in semi-supervised manner. • Affordable: To be able to ensure the wider availability, especially to allow penetration into developing countries, the cost of a single device must be affordable without compromising the quality and features. • Portable:

One of the main application is on-site detection, to make it

possible the solution need to be a battery powered handheld device. For advanced processing and computations, a mobile computer will be used which will essentially bring the cost of the device down. • Power consumption :

As it is battery powered, maximum power dissi-

pation is limited.

1.1

Intracranial Hematoma Detector (iHD)

This report describes the design, implementation, and evaluation of iHD system (figure: 1.1). It consists of hardware device - intracranial hematoma detector terminal (iHD terminal), firmware running in the terminal and software application intracranial hematoma detector mobile application (iHDMA). The device operates 3

Figure 1.1: iHD with its top enclosure removed.

in several modes and the fundamental operation is to transmit a burst of pulsed energy of Near-Infrared (NIR) light, let it take a banana shaped path [6] along the cranium (figure: 1.2), and measure the intensity on its reception.

Figure 1.2: The assumed banana-shaped light path through tissue sample.

Analog to digital converted value of intensity corresponds to the absorption of NIR light along its path. Using conventional Optical Density (OD) calculations presence of hematoma can be directly calculated [7]. However, to reduce the

4

complexity and improve the feature set, we propose few enhancements that are based on radar system principles. Instead of single scan, continuous scans will be carried out to generate several samples that will be put through various methods and pattern matching algorithms to semi-automatically detect the presence of hematoma, and provide visual imaging of the head.

1.2

Organization of the Report

The remainder of this report discusses the background, design and development of iHD system and evaluation of the system in real world application. Chapter 2 introduces to the general background and related work. Chapter 3 discusses operation of the terminal and mobile application. We present the implementation of the hardware and firmware of iHD system in Chapter 4. Chapter 5 describes the algorithms used for hematoma likelihood estimation in iHDMA. Chapter 6 presents the evaluation of the solution developed and discusses the results. Finally, Chapter 7 summarizes our work and discusses the possible future directions.

5

Chapter 2 LITERATURE SURVEY

Intracranial hematoma is one of the profoundly studied injuries in medical surgeries. Intracranial hematoma is a treatable cause of secondary injury which can cause significant disability or death if not promptly recognized and treated. They occur as the primary injury in 40% of patients with severe head injury. Recurrent hematomas, postoperative epidural hematomas, and delayed traumatic intracerebral hematomas develop in up to 23% of patients with severe head injury [7]. Near infrared spectroscopy (NIRS) is a spectroscopic method utilizing the near infrared region of the electromagnetic spectrum (from 700 nm to 1400 nm). NIRS can be used for non-invasive assessment of the brain function through an intact skull in human subjects by detecting changes in blood hemoglobin concentrations. This application is sometimes called optical topography (OT) in which NIRS is used for functional mapping of the human cortex. This chapter presents the background and discusses the related work in the area of intracranial hematoma detection. Section 2.1 begins with a general overview of intracranial hematoma and discusses the causes and implications. Section 2.2 lists the current diagnosis procedures, merits and demerits. Section 2.3 gives an overview of NIRS, examines its applicability for intracranial hematoma detection and surveys related work in this area. Section 2.4 discusses the safety regulations. Section 2.5 examines different techniques to estimate the presence of hematoma from intensity data.

6

2.1

Intracranial Hematoma (IH)

An intracranial hematoma is a hemorrhage that occurs within the skull. Intracranial bleeding occurs when a blood vessel within the skull ruptures or leaks. It can result from physical trauma or non-traumatic causes [8]. Intracranial hematoma possesses a serious medical emergency because the accumulation of blood within the skull can lead to increase in intracranial pressure, which can crush delicate brain tissues or limit its blood supply thus cause a secondary injury. Pressure increase can be lethal under certain circumstance where it could leave a potentially deadly brain herniation. Early identification prior to neurological deterioration, is the key to successful surgical treatment. This is currently accomplished by serial CT scan because it is the only reliable method currently available.

2.1.1

Types of Intracranial Hematoma

• Intra-axial Hematoma • Extra-axial Hematoma – Subgaleal Hematoma: Hematoma that occurs between the galea aponeurosis and skull periosteum. – Cephalhematoma: Hematoma that occurs between the skull periosteum and skull. – Epidural Hematoma: Hematoma that occurs between the skull and dura mater. The condition is potentially deadly because the buildup of blood may increase pressure in the intracranial space and compress delicate brain tissue. More often, a tear in the middle meningeal artery causes this type of hematoma. When hematoma occurs from laceration 7

of an artery, blood collection can cause rapid neurologic deterioration. Although it occurs in 1-3 % of head injuries between 15 and 20% of patients with epidural hematomas die of the injury [9] [10]. – Subdural hematoma: Hematoma that occurs between the dura mater and arachnoid mater. They often occur as a primary head injury due to fast changing velocities within skull, that leave tears in small bridging veins. Much more common injuries happen due to various rotational or linear forces. Further, it is more common in patients on anticoagulants, such as aspirin and warfarin, and can have a subdural hematoma with a minor injury. The associated mortality rate is high, approximately 6080%. Traditional methods like CT scan or MRI scan are commonly used to detect subdural hematomas [11]. While small subdural hematomas can be managed by careful monitoring until the body heals itself, larger or symptomatic hematomas require a craniotomy. It is also common on postoperative complications include increased intracranial pressure. The injured vessels must be repaired. – Subarachnoid hematoma: Hematoma that occurs between the arachnoid mater and pia mater (the subarachnoid space), is a form of seizure and a fatal medical emergency. It can lead to death or disability even when diagnosed and treated at an early stage [12]. Subarachnoid hemorrhage may occur in cases of TBI in a manner other than secondary to ruptured aneurysms, being caused instead by lacerations of the superficial micro vessels in the subarachnoid space [4]. If not associated with another brain pathology, this type of hemorrhage could be benign. It is a frequent occurrence in traumatic brain injury, and carries a poor prognosis if it is associated with deterioration in the level of consciousness. [13].

8

2.1.2

Traumatic Brain Injury (TBI)

The prime cause for the intracranial hemorrhage is Traumatic Brain Injury. TBI (figure: 2.1) is often a result of direct hit to the brain from an external mechanical force or high acceleration.

Figure 2.1: A detailed description of Traumatic Brain Injury.

The yearly incidence in the US is estimated to be 1.5 per 1000 people (1.3 mild, 0.15 moderate and 0.14 severe injuries) which contributes to 52,000 deaths annually. About two million people suffer from TBI and about 500,000 are hospitalized for TBI in the United States alone [11]. In developing countries such as Sri Lanka, the incidence of TBI has risen due to the alarming increase in automobile use and 9

industrialization with the absence of proportionate development in infrastructure, and therefore corresponding rise in the number of vehicle accidents. The mortality rate is estimated to be 21%, 30 days after TBI [11]. It is significant as large percentage of TBI deaths occur weeks after the event [13], mainly due to the secondary injury and complications developed. These secondary injuries exacerbate the damage and contributes to great number of TBI deaths occurring in hospitals [14]. Outcome for patients with head injury depends heavily on the cause. Patients with TBI from falls have an 89% of survival rate while only 9% of patients with firearm-related TBIs survive. Firearms and vehicle accidents are the most common cause of fatal TBI [11].

2.2

2.2.1

Other diagnosis procedures

Computed Axial Tomography (CAT)

CAT is the definitive tool for accurate diagnosis of an intracranial hemorrhage. The early detection of the aforementioned blood clots is paramount, and will be a life saver. Currently only Computer Tomography (CT scan) is capable of identifying it, nonetheless unfortunately, it is prohibitively expensive and rarely available. However, a clinical pre-screening technique could improve the utilization of CT scan.

2.2.2

Magnetic Resonance Imaging (MRI)

A medical imaging technique used to visualize the internals and functions of the organs and body. MRI offers greater contrast between different soft tissues and clear picture than CT does, and hence useful in brain imaging. Further, unlike 10

CT, it uses no ionizing radiation, but a powerful magnetic field and therefore less harmful to body. However MRI cannot be used on patients with metal implants, and cardiac pacemakers due to effects of the strong magnetic field and powerful radio frequency pulses [15]. Aforementioned methods share many merits like, high resolution, reliable operation and excellent accuracy. Nonetheless they also share the following common drawbacks • High capital cost and hence unaffordable in developing countries. • Complex operational procedures and hence require high operational, maintenance and repair cost. • Unportable and hence cannot be deployed with field units to be used for on-site detections. In addition, CT scan is highly harmful to the tissues as continuous high power irradiation could leave the patient with damaged cells thus increasing the likelihood of cancer and other complications.

2.3

Near Infrared Spectroscopy (NIRS)

Near-infrared spectrum is 0.75 − 1.4 µm in wavelength, defined by the water absorption. Near infrared spectroscopy is based on molecular overtone and combination vibrations. Such transitions are forbidden by the selection rules of quantum mechanics [16]. As a result, the molar absorption in the near IR region is typically quite small thus it penetrates much farther into a sample than mid infrared radiation. As a result, near-infrared light can penetrate several centimeters of biological tissues, enabling noninvasive investigation of the brain from the surface of the scalp. 11

Near-infrared spectroscopy (NIRS) is an optical noninvasive method of measuring cerebral hemoglobin distribution. It is a useful technique to investigate biological tissues, because in the near-infrared regions 750 − 900 nm (figure: 2.2), water has a low absorption, while oxyhemoglobin (HbO2 ) and deoxyhemoglobin (Hb) still have detectable absorption differences [12].

Figure 2.2: Absorption of light energy by Water and Hemoglobin in NIR region [6].

It is a relatively simple technique that is portable, does not require a dedicated technical staff, and does not require the patient to be injected with any isotopes.

2.3.1

Application of near infrared spectroscopy for intracranial hematoma detection

The basic principle of hematoma detection with NIRS is that water absorption in the near infrared range is relatively small and hemoglobin contributes to most of the tissue absorption; extra vascular blood absorbs NIR light more than normal brain tissue since there is a greater concentration of hemoglobin in an acute hematoma. By comparing the re-reflected and diffusing optical signal I2 from the 12

suspicious hematoma side and I1 from the healthy side or from a standard model, the optical density (OD) is calculated as in (Equation 2.1). OD = log10 (I1 /I2 )

(2.1)

The paper “Use of near infrared spectroscopy to identify traumatic intracranial hematomas” [7] claims that an OD value greater than 0.05 hints the presence of hematoma and likelihood increases with the OD value. An OD value greater than 0.5 definitely corresponds to an extra cerebral hematoma.

Figure 2.3: Histogram of Optical Density for different hematomas [17].

Figures 2.3 and 2.4 illustrate the direct correspondence between OD, and the presence and type of hematoma. Hemorrhages, irrespective of their type, depth or thickness produce an OD > 0.05 at high probability. Further extra cerebral hemorrhages, irrespective of the depth or thickness produce an OD > 0.6 [18]. Intracerebral hemorrhages fall in the gray region and may be confused with small, 13

Figure 2.4: Optical density variations under different hematoma conditions [17].

extracerebral hemorrhages as well. However as intra cerebral hemorrhage is hardly operable and tiny hemorrhages are not life-threatening, it can be referred to future analysis.

2.4

Safety Regulations

Food and Drug Administration (FDA), USA regulations limit the maximum power irradiated to the skin to 2−5 mW/mm2 in the spectrum of wavelength of 700−900 nm for the exposure of 10 s or more. An LED source of high light intensity can be considered for use when faced with this limitation. Based on the experiments by Ito et al. [19], elevations in the temperature due to NIR absorption is less than 0.5o C. In the experiments by Alper Bozkurt, et el [20], the increase in the skin temperature was 0.5 ± 0.1o C due to the cushioning 14

material used to attach the light source, and that was mainly due to skin-air heat exchange and sweating. However, the temperature increase due to semiconductor junction is in the range of 1 to 10o C [20]. The major contributor is therefore the semiconductor junction, which may cause burn injuries [21][22]. Further cell death is possible in the event of sustained temperature rise above 41o C [23]. Therefore care must be taken to control temperature. It is required to control irradiation due to LED to 50 mW/cm2 . This is comparable to the irradiance of the NIR region of the sunlight, which is about 50 mW/cm2 . The temperature increase and total irradiated power can be controlled by pulsing the current that is fed to NIR LED source.

2.5 2.5.1

Algorithms Threshold detection

In radar systems, to detect the hit, dynamic threshold is preferred to constant threshold, because noise level could vary. This ensures the constant probability of false alarm, while varying the threshold level accordingly. When standard deviation (σ) and average noise (Vn ) can be calculated from the sample vector, the dynamic threshold is established as in equation:2.2.

Vm = Vn + k × σ

2.5.2

(2.2)

Post detection integration

Similar to the aforementioned algorithm, post detection integration is a popular method used in radar systems. Here, once positive indication vector is generated, unusual pattern that highlights an impossible situation could be identified. 15

In mathematical terms the probability estimation can be given by equation: 2.3, for n samples with npositive number of positive indications. Positive indications with less probabilities can be thus filtered out.

ppostdetectionintegration =

npositive n

(2.3)

Complex pattern matching algorithms can be used, and we demonstrate the application of bidirectional associative memory in chapter 5.

2.5.3

Bidirectional Associative Memory (BAM)

BAM is an important tool available in Neural Networks and one of the most widely used. It achieves heteroassociation with smaller correlation matrix. Once trained, using a set of pairs of vectors, it could recall a pair (A, B) given initial pair (α, β). Wang et al [24] illustrated the use of BAM for pattern recognition. Also it was proved by Kosko [25], that BAM will converge for any correlation matrix W. However it could not guarantee that energy will be at local minimum. But it was shown by Yeou-Fang Wang et al [26] that guaranteed recall exists for all training pairs thus trained relationship could be established. Additional mathematical derivation is provided in Appendix D.

2.6

Chapter Summary

This chapter described background studies. Also it presented the current approaches and related work. The chapter then continued to explain the application of NIR in IH detection. After that, it proposed enhancements in detection algorithms to provide a new feature set and to improve accuracy. The next chapter discusses the operation of iHD system. 16

Chapter 3 SYSTEM OPERATION

The complete solution consists of a portable remote device - iHD terminal - and a mobile application - iHDMA - running on a mobile computer such as a Personal Digital Assistant (PDA), notebook or netbook. In this chapter we will look at the use of the device and operation of both iHD terminal and iHDMA. This chapter is provided for the completeness only, and for a beginner level instruction User-Guide must be referred to. Generic overview is given in section 3.1. Section 3.2 discusses the use of iHD in-depth and section 3.2 leads through the use of iHDMA.

3.1

Using the system

The iHD terminal is used to perform the scans and the iHDMA is used to compute the results and visualize them.

• Mode 1 - Active Scan :

Continuous scan over the injured side of the

head to obtain the intensity values over scanned area. • Mode 2 - Reference scan : Continuous scan over the healthy side of the head to obtain reference values. • Mode 3 - Quick scan : Quick scan is performed in four stages over four predefined locations (figure: 3.2) on the head. On each location a pair of scan is performed on each side starting with left side of the head. Realtime Bluetooth communication is enabled in this method. 17

Figure 3.1: Terminal is operated in four different modes. User can sequentially move from one mode to another mode by pressing the mode button on the device. User is also allowed to skip through modes and come back later.

18

– Mode 3a - Quick Scan on Frontal area : Over left (followed by right) side of the forehead above the frontal sinus – Mode 3b - Quick Scan on Temporal area : On left (followed by right) side of the temporal fossa – Mode 3c - Quick Scan on Parietal area : Above left (followed by right) side of the head midway between the ear and the middle of the skull. – Mode 3d - Quick Scan on Occipital area : Behind left (followed by right) side of the head midway between the ear and occipital protuberance. • Mode 4 - Data Transfer : iHD terminal’s Bluetooth module is put in the listening mode so that a mobile application can query data from the iHD terminal. Further this mode allows few low level firmware configurations.

Figure 3.2: Although Optical Density scans can be taken at any two identical positions on either side of the head, these predefined positions will make the process simpler and efficient.

Once the scans are completed, the data is transmitted to a mobile computer running iHDMA for analysis and semi-automatic diagnosis. iHDMA is run on a .Net platform powered mobile computer with Windows Operating System (OS) installed. The computer should be Bluetooth enabled. The estimates are computed by iHDMA and visualized (figures:3.3-3.4).

19

Figure 3.3: An important objective of iHD is to be simple to operate. The results are visualized through a brain imaging, and thus provides greater usability.

Figure 3.4: The summary is provided for a quick glance of all the details.

20

3.2

Operation of iHD terminal

Once the iHD terminal is powered on it will be put in mode 1 - active scan. The user is indicated that the device is ready and waiting for user input. The device is then moved gently over the injured side of the head from forehead to the occipital protuberance while keeping the scan button pressed. The active state is indicated to user. An average movement of 160 mm is expected at 16 mms−1 .On average this will yield more than 400 samples. If required, it could be performed over shorter length at a slower speed resulting in concentrated samples for higher resolution. It is followed by mode 2 - reference scan where identical procedure is repeated but over the reference side of the head. Should the scan be made over a shorter length during active scan, it must however be matched here. Although approximately same number of samples are expected the exact number could vary and it is considered in the algorithm. If a scan is repeated under a particular mode before uploading the results to the mobile application, data will be replaced allowing repeated scans to correct errors in the process. To support conventional OD measurements and facilitate the comparisons, mode 3 (3a to 3d) - Quick scan is provided. Under this mode, the device is kept over the indicated position on the left side of the head and scan is carried out by pressing the scan button once. It must be followed by the same procedure over the same location on the right side of the head. For example under mode 3a, firstly the device is placed on the left side of the forehead above the frontal sinus and scanned. Then it is repeated on the right side of the forehead above the frontal sinus. Realtime Bluetooth connection is enabled under this mode. Similar to mode 1 and 2 repeated scans under a given mode will replace the previous data.

21

Once the required scans are completed, it moves into mode 4 - data transfer. Here, the device is put into discovery mode searching for a serial profile enabled Bluetooth connection. If one is found it is added to the list . No more user interaction is needed on device side to complete the data transfer or diagnosis.

3.3

Operation of iHDMA

The use of mobile application is intuitive and designed to offer optimal user experience while enabling the maximum flexibility. The application is written on .Net 3.5 platform1 and has no other dependency. The standard flow is to begin with data capture (figure: 3.5). Under this step, user could verify the presence of a valid iHD terminal by pressing connect button. The device would respond with a code that will be used to identify a valid terminal. User could then proceed to capture data.

Figure 3.5: Under capture mode following the device validation user could import the data with a single click. Should any error occur data buffer could be flushed with reset button.

Capture is followed by save option, where record ID, medical officer ID and other details can be stored together with the data. Also the integrity of the data is 1

.Net 3.5 runtime can be http://www.microsoft.com/downloads

downloaded

22

from

Microsoft

website

:

checked and it is processed (figure: 3.6). The data is stored in XML format so that it could be subjected to further analysis with external applications.

Figure 3.6: Captured data can be completed with additional parameters and saved as an XML file. User can then proceed to process data to generate analysis. Also user could load already saved or processed data for a re-analysis. Further availability of active scan, reference scan and OD scan data are indicated to the user; so is the data analysis results.

In the next step a quick snapshot of the results is produced for instant diagnosis. Both intensity threshold (based on active and reference scans) and conventional OD results are used to create visual imaging. The color variation is used to indicate the severity and when the threshold is exceeded alert is generated, thus enabling the quick identification of a hemorrhage. It is followed by advanced analysis (figure: 3.7). The last step offers advanced configuration (figure: 3.8) for experienced users. If Bluetooth connection port is configured with non-default parameters, corresponding settings can be made here. Further signal intensity, pulse frequency and other related parameters of iHD system can be configured here and synchronized with the device.

23

Figure 3.7: In addition to quick results, advanced analysis can be opted for. Parameters like normalization count, threshold limit, etc can be set with the aid of visual indication under this section. Hence the parameters can be set to attain the optimal performance.

Figure 3.8: Mobile computers may offer Bluetooth through a different port or with different settings. Further user might prefer to change iHD terminal’s firmware parameters for better performance. This section permit such configurations over Bluetooth, without modifying the actual firmware code. Also it allows global level iHDMA parameter configurations.

24

3.4

Chapter Summary

This chapter described the use of iHD system. This chapter also described the detailed use of software (iHDMA) and hardware (iHD) components of iHD system. The next chapter discusses the architecture design and implementation of iHD system.

25

Chapter 4 SYSTEM ARCHITECTURE

The critical component of the solution is the iHD terminal. In this chapter we will look at the design architecture of the iHD terminal. We begin the chapter with detailed description of the architecture of the iHD terminal in Section 4.1. We look at the system from bird’s eye view and analyze the inter-operation between the components. Section 4.2 is dedicated to the hardware design and implementation. In Section 4.3 we describe the firmware component running the iHD terminal. Section 4.4 examines the communication protocol specifications.

4.1

iHD System Architecture

iHD terminal is a small hardware platform consisting of a microcontroller, NIR transmitter and receivers, and other associated hardware units. Detailed explanation is given in section 4.2. The unit operates in four different modes. Detailed instruction on operating the device is given in chapter 3. iHD terminal is used to perform different types of scans over both injured side and reference side of the head, store them and transmit to the remote application. The fundamental operation of a scan is to transmit a pulse modulated signal and measure the intensity of the received signal. The transmitted signal is confined to NIR spectrum with the aid of NIR LEDS, and differential analysis is performed to cancel ambient noise. This atomic step is repeated at configured frequency on

26

a predetermined pattern to achieve several modes of operation. Further parameters pertinent to the signal generated and reception component are maintained configurable to enable maximum flexibility.

4.2

iHD Hardware Architecture

Figure 4.1: iHD Unit.

The hardware implementation of the system (figure: 4.2) can be grouped into five basic modules each with unique functionality considering the overall architecture. The basic modules are, • Signal generation and Sensor Subsystem • Processing & Controlling Subsystem • Communication Subsystem 27

• Power Management Subsystem • Input Output Subsystem

Figure 4.2: Architecture diagram depicts the interaction between individual subsystems.

4.2.1

Signal generation and Sensor Subsystem

Concentration of hemoglobin determines the extent of absorption of NIR signal and thus affects intensity of reception. Hemoglobin concentration in a particular point of the brain is found relatively by transmitting series of pulses of Near Infra Red light and measure the reflecting amount of NIR energy, through the NIR intensity to voltage convertor reading. The amount of reflected energy is a measure of amount of absorbed energy by brain tissues. The process can be further divided into the following • Pulse generation and Current Control • NIR light emission 28

• Signal reception, noise cancelation and analog to digital conversion • Interfacing, and Transmitter - Receiver Separation

4.2.1.1 Pulse generation

Burst of pulses are generated through Pulse Width Modulation (PWM) programmatically. This facilitates the power and time control in software. Microcontroller is used to generate the PWM signal (figure: 4.3). Although it was initially considered to use an isolated signal generation circuit for better performance, due to above reasonable power line noise, microcontroller based signal generation was chosen. Further it allows greater control over the frequency, duty cycle and burst time, while maintaining them configurable through firmware settings. Further external transistor based amplifier is employed to permit larger current and it acts as a current source. To prevent over current condition, PIC is coupled through inverters that act as buffers.

Figure 4.3: Burst of pulses that are fed to NIR LED.

4.2.1.2 NIR light emission

The wavelengths selection is based on the absorption spectra of the hemoglobin. As explained earlier peak difference in absorption between water and blood is at 760nm for Hb and 850 − 860nm for HbO2 . 29

Wavelength consideration • 805nm - 810nm : A wavelength of 805nm - 810nm is suitable for hematoma detection since it is very close to the isosbestic wavelength of oxyhemoglobin and deoxyhemoglobin absorption, and the signal detected will not be affected by differences in oxygen saturation in blood. Narrow bandwidth high power NIR LEDs within this frequency range, that were purchased from ROITHNER LASER TECHNIK are used in iHD. • 760nm - 850nm : Wavelengths of 760 nm and 850 nm are selected to monitor temporal changes of cerebral concentrations of HbO2 and Hb. For each wavelength it is assumed that the linear changes in attenuation for each chromophore can be linearly summed. The result of these computations is the value of the absolute change in concentration of each chromophore in the non-arbitrary units of micro molar of chromophore per liter of tissue. Two wavelengths can be multiplexed to have separate information on each wavelength. They can be alternatively turned on and off in pulsed manner to achieve this. As the latter approach clearly possess an advantage it was attempted initially. However due to the fact that the sensor has equal response in both the bandwidths, mutual noise became critical. It was however later handled by introducing an optical narrow bandwidth filter at 760nm and 850nm. This produced significantly better results. Nonetheless, due to the lens attenuation the intensity reception did not yield results better than single wavelength operation, and thus considering the simplicity of the design, peak difference in absorption levels between water and hemoglobin, and ambient noise, 850 nm single wavelength emission is used in iHD terminal. One of the most important reasons for choosing a LED is that its radiation is within the FDA limitations for the radiation power, and if needed multiple LEDs can be 30

operated together to attain required radiation intensity. Further glass top convergence provided by LED was adequate to achieve the radiation intensity required to penetrate the tissues and skull. Compared with quantitative measurement of oxygen saturation or other applications, the incident light for hematoma detection does not require as sharp a spectrum distribution as does the laser; hence LED is used.

4.2.1.3 Signal reception, amplification and Analog to Digital Conversion (ADC)

Following types of detectors, that can be used to measure the transmitted signals, were considered. • Silicon photodiodes (eg OPT101) • Avalanche photodiodes (APD) • Photomultiplier tubes (PMT) • Charge coupled devices (CCD) The latter two options are not feasible on the basis of price and size. Silicon Photodiodes (SiPD) : Although silicon photodiodes have a lower sensitivity in comparison to APDs, their reasonable response time, high dynamic range, and the requirement that several needed to be placed within the flexible probe area, made them the perfect choice. The integrated combination of photodiode and transimpedance amplifier on a single chip (figure: 4.4) eliminates the problems commonly encountered in discrete designs such as leakage current errors, noise pick-up, and gain peaking due to stray capacitance. OPT101 is a monolithic photodiode with on-chip transimpedance amplifier (figure: 4.5) from Texas Instrument (TI) which is specially made for medical applications. 31

Figure 4.4: OPT101. An integrated solution for high sensitive light to voltage conversion over NIR region.

It is inexpensive compared to others such Si sensors and costs only 8U SDs [27]. TI’s free sample offer was used for the research and development process.

Figure 4.5: The 0.09×0.09cm2 photodiode is operated in the photoconductive mode for excellent linearity and low dark current. The OPT101 has high sensitivity of 0.45A/W and quiescent current is only 120A. Further peak response is at around 850 nm [27].

Dark Current Offset Correction:

The dark current is the result of absence

of light falling on the photodiode and it makes small voltage at the output and it should be avoided to have better performance. The photodiode dark current of OPT101 is approximately 2.5 pA and contributes virtually no offset error at room temperature. The bias current of the op amp’s 32

summing junction (- input) is approximately 165 pA. Further it can is deducted with the following circuit for improved performance (figure: 4.6).

Figure 4.6: The dark current will be subtracted from the amplifier’s bias current, and this residual current will flow through the feedback resistor creating an offset. The dark output voltage can be trimmed to zero with this optional circuit. [27].

4.2.1.4 Interfacing and Transmitter - Receiver Separation

The primary concern was to achieve maximum coupling between transmitter and receiver along the path through cerebrum while ensuring safety, minimum interface noise addition, and least interference from ambient noise. Separation : It was noted during the tests, due to the high sensitivity of the sensor and narrow bandwidth of the transmitter, the direct noise from transmitter to sensor could prevail at a meter separation even when they are not facing each other. Further it was noted that the received power increases with the separation up to 5 cm and then decreases according to inverse square law. The objective of the design is to ensure maximum coupling through cerebrum and thus the perfect separation was deemed to be 3 − 3.5 cm. 33

Interface : The photon current will be significant and unstable at the contact boundary between the optical probe and the tissue surface. Poor optical contact results in noise, false signals, and inconsistent readings. The surface of the head is usually not flat. Further it was adequate to maintain perfect contact with sensor while ensuring the emitted power of the LED is tunneled through the skull. It is achieved by raising the sensor slightly above the surface of the device while placing a pair of LEDs in a pit with aluminum and black rubber. This guarantees that full LED power is channeled through head at minimum direct leakage to the sensor (figure:4.8).

Figure 4.7: The design is planned such that, if necessary this could be further improved to accommodate flexible elastic optical probes as shown.

The flexible probe improves optical contact, but we cannot say that it totally solves the contact problem; for example, movement of the skin also results in signal instability [28]. Thus the movement needs to be gentle and we advice a speed of 16mms−1 . 34

Figure 4.8: The LED and photodiode are mounted on the baseboard, which can be curved a little to fit the shape of a human head. A black sponge is used to prevent leakage of light from the source to the detector [28].

Encapsulation : LED cannot be placed in direct contact with skin as the heat transferred through conduction will be significant and will result in high temperature rise of the skin (up to 9 degrees). This is taken care of by placing the LED in an abyss while ensuring total enclosure. It requires encapsulation in a suitable material: • Flexible enough to conform to the head • Have suitable optical properties • Do not give off any by-products A two-part clear silicon rubber satisfied all of the above requirements. Further, to prevent direct leakage, it was first wrapped by aluminum foil followed by the rubber material coated in black color. Encapsulation was carried out in three stages • Encapsulation of the back of the sensor board • Encapsulating the front of the PCB of LED board, except for the regions directly above the LEDs and photo sensors 35

• Filling the regions directly above the LEDs and photo sensors

4.2.2

Processing & Controlling Subsystem

This is the brain of the blood clot detector and responsible for controlling all the subsystems. The microcontroller is in charge of the whole operation, which includes acquiring the signal, adjusting the incident light intensity, and communicating with the operator. The main functionalities required for this subsystem will be Analog to Digital Convertor (ADC) to interface Sensor, controlling NIR LED currents, adjusting intensity via use of pulse circuit, and UART operation for serial communication between Bluetooth and Microcontroller. Due to the fact that whole system is operated by low capacity small battery, it should operate with low power consumption.

4.2.2.1 PIC Microcontroller

The major reason for selection of the PIC microcontroller is that it has emerging development tools. Other reason for this selection was, this project was selected to the second phase of the Microchip PIC32 Design Challenge. We were offered the starter kits and other required resources free. Hence we selected PIC family of Microcontroller as our main processing and controlling unit. Features of PIC18F452 • 2 level priority interrupts • 10 MHz Maximum Frequency • 32K Program Memory • Multiple Power Management Modes 36

• UART Module • 8-Channel 10-bit Analog-to-Digital Converter

4.2.2.2 Sensor Interfacing

The OPT101 sensor produces an analog output which varies among two voltage levels. But microcontroller processes only digital data. So we need to use analog to digital conversion when interfacing the sensor to the Microcontroller. The microcontroller has a 10 bit Analog-to-Digital converter (ADC) with externally configurable voltage references. The 10-bit ADC includes the following features: • Successive Approximation Register (SAR) conversion • Up to 500 kilo samples per second (ksps) conversion speed • External voltage reference input pins • One unipolar, differential Sample-and-Hold Amplifier. The data available from the sensor input can be read as digital after converting them from analog by using ADC of the microcontroller. The absorption of NIR light is the difference of the transmitted energy to the received energy. But these two parameters cannot be calculated in practice. The basic operation of hematoma detection is by exploiting the fact that water absorption in the near infrared range is relatively small and hemoglobin contributes to most of the tissue absorption which is calculated by comparing the reflected and diffusing optical signal intensity Ilef t from left side and Iright from right side, the optical density OD can be derived to: OD = log10 (I0 /Ilef t ) − log10 (I0 /Iright ) = log10 (Ilef t /Iright ) 37

(4.1)

Then intracranial hematoma can be detected by comparing the OD with the detection threshold and historical data. According to position, different intracranial hematomas can be divided into epidural, subdural, and intra-cerebral types. Background study given in section 2.3 provides the insight into this.

4.2.2.3 Possible methods of computation • Time resolved : In Time-resolved method, short pulses of light is transmitted and the distribution of time of flight of the transmitted photons is measured. This measurement provides the greatest amount of information about the tissue being investigated, but the device will be complex. • Frequency modulated : Frequency modulated instruments involve modulating the light source at radio frequencies and detecting the intensity and phase of the transmitted signal. This requires less complex instrument. • Continuous Intensity : The continuous intensity instrument is the simplest of all, where light is injected into the tissue and the attenuated transmitted intensity is measured at some distance from the source.

The exact depth of the hematoma from the surface that can be examined by NIRS is still controversial. Further direct depth calculation is rendered impossible due to short distances involved. The typical time of flight for a depth of 3cm will be Tf light = (2 × 0.03)/(3 × 108 ) = 2 × 10−10 s

(4.2)

We will neither be able to generate nor detect at this high frequency (Equation : 4.2) in a microcontroller based implementation. However through transmitting

38

burst of pulses of continuous intensity (type 3) and detecting patterns the characteristics of the hematoma can be found.

4.2.3

Communication Subsystem

Figure 4.9: GS-BT2416C2 Bluetooth Module

The use of a remote connection to a wireless device such as PDA can help us integrate great variety of flexibility into the system. The mobility of the device is extremely high in such a design. Also it increases user friendliness by use of graphical user interface on the remote device. Reasons for use of a Bluetooth (BT) module (figure: 4.9): • Cost effective comparing to high-end microcontrollers and graphical display on the device itself • Easy integration with microcontroller via Universal Asynchronous Receiver Transmitter (UART)

39

• Relatively flexible, feature-rich software development and better graphic quality are achieved. With use of BT we can maintain our functionality in better quality and in very economic manner than other options. A suitable BT module called GS-BT2416C2 is used as it possess the following characteristics. 1. SPP (Serial Port Profile) Support : All Mobile phones/PDAs support SPP only 2. Direct interfacing to a UART and control by AT commands 3. After initialization with AT commands it direct data transfer via UART 4. Bluetooth specification V.1.2 compliant 5. Transmission rate up to 721 Kbps 6. Working distance up to 10 meters without an external antenna 7. Hardware based UART flow control The BT module is compatible with UART interface controlling (figure: 4.10). The module with AT command is dedicated to implement serial cable replacement. An automatic point to point connection takes place when modules are switched on. Modules are configured via macro instruction to play the role of master or slave.

4.2.3.1 UART

UART is used to control the module with AT commands (COMMAND MODE), or send/receive serial data to be transmitted over the SPP Bluetooth link (DATA MODE). The first time that the module is powered up, the default UART settings are as following (saved in flash memory) 40

Figure 4.10: The interfacing of the BT module using the UART and 2 General Purpose Input Output(GPIO) ports.

1. Baud rate: (bps) 9600 2. Data bits: 8 3. Stop bits: 1 4. Parity: None 5. Flow control: None When these settings are changed by the AT+UARTSETUP command, they are stored in the flash memory to be reloaded when the module is powered up the next time. GPIO1 : This will be configured as output. GPIO1 is high when an SPP Bluetooth link to a remote device is present. GPIO1 is low when no Bluetooth link is present. GPIO3 : The GPIO3 should be configured as input. If GPIO3 is set to high, the module switches its mode of operation to DATA MODE. If GPIO3 is set to low, the module switches its mode of operation to COMMAND MODE. After aforementioned initialization is performed, any feature of the BT module can be 41

accessed via the its protocol stack (figure: 4.11). The dataflow and command sets are explained under section 4.3.

Figure 4.11: Bluetooth Stack

4.2.4

Power Management Subsystem

This subsystem manages power to the rest of the subsystems. To maximize the power capacity and portability a rechargeable dual cell (7.4V - 8.4V ) will be used to power the terminal. Provision of power at the appropriate voltages to the other subsystems will be handled by this subsystem. The voltage and current requirements for the device is tabulated below (Table 4.1). Device / Subsystem PIC18F452 Bluetooth Module Sensor Subsystem I/O

Vmin (V)Vtypical(V)Vmax (V)Imax (mA) 2.3 3.3 3.6 200 3.13 3.3 3.47 90 2.7 5 6 250 3 5 6 50

Table 4.1: voltage and current requirements of iHD terminal subsystems.

4.2.4.1 Low Dropout (LDO) Regulators

The REG1117 is a family of easy-to-use three-terminal voltage regulators from Texas Instrument. The family includes a variety of fixed and adjustable-voltage 42

versions, two currents 800mA and 1A. REG1117 characteristics

• Output tolerance: 1% • Output Current: 800mA, 1A • Dropout Voltage: 1.2V @ 800mA • Internal Current Limit • Thermal Overload Protection • 0.06V voltage ripple at 3.3V Output • 0.10V voltage ripple at 5.0V Output

4.2.4.2 Power Source

The requirement to have portability of the terminal requires small size battery operation when considering the weight and the size of the device. Therefore we have selected rechargeable dual cell with capacity of 2100mAh for major power source for our device. The 2100mAh of capacity can operate the device continuously for 3 hours at full functionality.

4.2.4.3 DC Power and Charging

The device can be simultaneously powered by both battery and external DC supply. A generic purpose DC adapter with 9V , 1A rating could be used to charge the batteries while powering up the terminal. When the device is discharged the user can connect to the charger and keep until the batteries get fully charged. However to minimize the complexity and size of the device, no special power management unit is provided. 43

4.2.4.4 Sleep, Power Saving Modes

The iHD terminal is an on-demand-use device and hence not expected to be powered on when it is not in use. Therefore special power saving and sleeping modes other than the micro controller’s built in power saving modes of operation are not considered.

4.2.5

Input Output Subsystem

This is the subsystem that interacts with user. The Sensor Subsystem could also be categorized under this subsystem, but as it is the most critical component it was considered separately. In this section we focus on controlling inputs to the systems. Considering the portability, power and simplicity requirements, we moved the responsibility of processing and displaying results to the mobile computer and therefore iHD terminal is only providing minimal user interaction. The main inputs used here to interact with user, are the mode selection button and scan button. Visual imaging, statistical analysis and learning system will be implemented in remote mobile computer such as a PDA which will act on the data received via the Bluetooth connection. As explained in section 3.1, once the device is powered on it will be put in mode 1. Through a push button, user could navigate through modes. The button is connected to an interrupt enabled input of the microcontroller to give the highest priority. User can then trigger the scan by pressing and holding the scan button (under mode 1 - active scan and mode -2 reference scan) or pressing and releasing the scan button. The button is provided with de-bouncing algorithm to improve user interaction and allow accidental releases.

44

4.2.6

Safety concerns

As explained in section:2.5, care must be taken to ensure that the irradiated power is within the safety limits. It was also earlier established that radiation up to 50 mW/cm2 is harmless as it is comparable to the irradiance of the NIR region of the sunlight. The average radiated power of an NIR LED is 20 mW, thus resulting in total radiation of 40 mW assuming that 1 cm2 area is illuminated.

4.3

iHD Firmware Architecture

Figure 4.12: iHD firmware block diagram

iHD Firmware Architecture is shown in (figure: 4.12). Due to the realtime nature of the application both external input and internal clock triggered interrupts based scheduler is implemented to switch between different states. A user input will take priority and hence result in the state change wherever the program was. However necessary care is taken to windup the current state properly and variables are properly stored.

45

State maintainer chooses the appropriate state function and runs. There are two such state implementations. In addition, it handles the user indication output that reflects the current state and status of the operation.

4.3.1

Communication state implementation

This module handles the Bluetooth communication. When the state is entered, the component resets the BT module through hardware reset, thus escaping from whatever status the BT module is currently in. It is followed by initialization of AT commands, and initial settings. Then BT is put in discovery mode, and when it is contacted by a mobile computer running iHDMA it answers. This module is also responsible for the responses to the commands sent from iHDMA. In this context, it handles the formatting of replies and queueing them.

4.3.2

Data Acquisition state implementation

This module handles the different modes of scans. When one of the scan mode is entered, settings of an atomic scan operation is loaded from EEPROM and stored in global variables. Then the module chooses the properties depending on the scan mode and calls the atomic scan function that is common for all. The operation is then delegated to PWM generation, ADC sampling, encoding (10 bits output to a word) and store functions.

4.3.3

User Indication IO

The current state is displayed to user at all time, and changes in states are reflected instantly, irrespective of the unloading and loading latency involved. Further, the

46

operation status such as scanning, busy, ready to scan and errors are displayed appropriately.

4.4

Communication Protocol

This section describes the protocol that is adopted to be used for the communication between the remote intracranial hematoma detector (iHD) and Medical Application (iHDMA) that is running in a Personal Computer, Notebook or Mobile Computer. Detailed description of the protocol and examples are given in Appendix B.

4.4.1

Introduction

The iHDMA will always be a host and iHD always be slaves. Therefore, only iHDMA can initiate communication by querying the iHD and the iHD is obliged to respond. The atomic unit of the data is a word (16 bits) which is in big-endian format, i.e. the higher order byte is sent first. All the messages are encapsulated between a message-starting header and message-ending header. Further frame headers wrap all data frames. One cycle of communication includes the command initiated by iHDMA and the reply from iHD. Commands can be one of the following : • Request Settings • Update Settings • Request Data • Ping 47

4.4.2

Structure

As aforementioned, a command or reply message is wrapped by headers. The body length can be of variable number of words. The header identifies the type of the word (i.e. a Message Header) and type of the command (Ping, Request Data, Request Settings or Update settings), and the body is supplemented with any parameters if available. Hence, the body may be absent when it is neither required by the command nor available. The command or reply is ended with message-ending header which is 0xF F F F for all messages.

4.4.3

Message Starting Header

A 16 bits long word of the following format (Table : 4.2) will serve as a messagestarting header. 1

1 1 1 Word Type (Message Header)

0 0 0 0 0 0 Reserved (0 always)

0 0 Command / Reply

X X X X Message ID

Table 4.2: Message Starting Header

There are four types of commands (Table : 4.3) and they are replied with the reply bit turned on: Ping Request Data Request Settings Update Settings

1 1 1 1

1 1 1 1

1 1 1 1

1 1 1 1

0 0 0 0

0 0 0 0

0 0 0 0

0 0 0 0

0 0 0 0

0 0 0 0

0 0 0 0

Table 4.3: Types of message commands

48

0 0 0 0

0 0 0 1

0 0 1 0

0 1 0 0

1 0 0 0

4.4.4

Message Body

For a command, message body is used to pass relevant parameters. For a reply, message body consists of information that is queried through the command and can vary depending on the type of the command.

4.4.5

Message-ending Header

All 16 bits are set to 1s and thus can be easily identified. Hence, data packets of any type, settings-replies, frame headers and ping-replies cannot have all 16 bits set.

4.4.6

Frame

A frame can be either a frame-idle or data-frame, again consists of frame-starting header, data packet and frame-ending header. Scan data is sent inside a frame when requested.

4.5

Summary

This chapter described hardware architecture of iHD terminal explaining the individual subsystems and their operations. It discussed the firmware architecture of iHD terminal. Also it presented the communication protocol specification. The next chapter discusses the algorithms used to estimate the likelihood of hematoma in iHDMA.

49

Chapter 5 HEMATOMA LIKELIHOOD ESTIMATION IN IHDMA

Traditionally, for optical density calculations, scans are performed at selected pair of locations, and they are used to estimate the likelihood of hematoma. We explore alternative algorithms, mainly derived from a completely different area - radar systems, for this purpose. In addition, we try to develop Artificial Neural Network (ANN) based model for semi-automatic detection. An introduction into optical density calculations is given in section 5.1 with a short discussion on its merits and demerits. Section 5.2 examines the characteristics of the intensity sample values. In section 5.3 we generate one-to-one reference estimates for each scan. Section 5.4 walks through threshold detection based estimation followed by post detection integration in section 5.5. Section 5.6 discusses the possibilities of using complex pattern matching algorithms and explains application of Bi-directional Associative Memory (BAM) based ANN model.

5.1

Optical Density calculation on Isolated Scans

As explained in section:3, mode-3 corresponds to quick scans on isolated predefined locations on either side of the head. The scans are performed in pairs starting from the point on the left side of the head followed by the measurement at the corresponding point on right side. It leaves us with four pairs of intensity measurements. Using equation:4.1, we could estimate OD for the given point.

50

Absolute value of OD tells us the likelihood of a hematoma. The background study given in section:2.5, provides us with a method of ascertaining it. The sign of OD gives us the hematoma side. A positive OD means right side’s intensity value is lower, and it translates to higher absorption of NIR on that side. Hence it could be inferred that hematoma is on the right side. Scans are performed on four predefined locations namely, frontal area, temporal area, parietal area and occipital area. When a positive indication is generated more scans can be carried out in the proximity and a clear picture can be generated.

5.1.1

Merits and Demerits

The merits lie in the simplicity of the method and quick analysis. It gives the results instantly, and an expert could offer the opinion instantly. Using his past experience, he could be able to make the estimations without further analysis. The demerits are possible false alarms due to noise and glitches that are not accounted for. Further the absence of visual indication, advanced analysis and intelligence makes this method unsuitable for less experienced personnel such as first-aid crew.

5.2

Characteristics of Intensity values and OD

Hematomas are generally of considerable size and hence identifiable from several scans close by. Sensor, transmitter separation of 3 cm browses through an area of 1cm2 and hence by taking several measurements around that point false alarms and misses can be avoided. With regard to the absolute value of OD, an OD > 0.5 corresponds to 3 times higher absorption and thus confirms the presence of hematoma. However an OD = 51

0.1 corresponds 1.26 times higher absorption which could have been caused by a glitch or noise. An expert opinion is hence required under this method. Through ANN we propose and automatic detection in the absence of an expert. Further, the ability to make continuous scans give a promise for visualizing the results easily. This requires us to acquire a number of samples over the injured side of the head and find the corresponding reference value.

5.3

Reference Estimation through normalization

The caveat in carrying out continuous scans is the difficulty in controlling the number of samples and uniformity. We could advice the personnel to move at a constant slow pace over the head thus achieving reasonable level of uniformity in the measurements. Further, knowing the total number of samples and end points on the head we could estimate the average length covered by each sample. However the problem of unequal number of samples from active scan and reference scans still exists. We handle this by normalizing the reference sample count to that of active scan.

5.3.1

Moving average based normalization

For N number of active scan samples and M number of reference scan samples, our objective is to achieve N number of reference scan samples (figure: 5.1 and 5.2) thus obtaining one-to-one correspondence between active scan samples and reference scan samples. This is done by the normalization function (equation: 5.1).

52

Figure 5.1: The image depicts a possible scenario where unequal number of samples might be acquired in Active and Reference scans.

Figure 5.2: The image illustrates the normalization operation. While unequal sample counts have been normalized, spikes in reference samples also have been been removed. This smoothing effect is controlled by the Normalization Factor and in realtime controlled by the slider shown in the diagram.

53

let, X be the active scan sample vector, Y be the reference scan sample vector, and Z be the normalize reference scan sample vector. f or1 ≤ i ≤ N, j = bi × N/M c   j − L if j − L > 1 Ns =  1 if else   j + L if j + L < M Ne =  M if else PN e

Y [k] Ne − Ns k=N s

Z[i] =

(5.1)

However, with the distance from the current sample, the influence of other sample values must decrease. That is if there are 20 active scan samples and 40 reference scan samples the normalized reference scan sample corresponding to the 12th active scan sample must heavily depend on 24th reference scan sample than on 26th or 22nd samples. This is achieved by a windowing function, for example, a linear windowing function would be as in (Equation: 5.2).      W1 =

   

k−N s j−N s

if N s ≤ k ≤ j

N e−k N e−j

if j ≤ k ≤ N e

(5.2)

0 if else PN e

Z[i] =

Y [k] × W1 Ne − Ns

k=N s

(5.3)

However, for a larger normalization count along with a higher number of samples, we go for an inverted-square windowing function (Equation: 5.4), that takes the 54

shape of

1 . 1+x2

  W2 =

5.4

1 1+(k−j)2



if N s ≤ k ≤ N e

0 if else

(5.4)

Threshold detection

Once we generate two vectors of equal size, they can be directly compared and anomalies can be identified. However to standardize the process and automate we have to create a mathematical model. The caveat is the intensity variation due to absorption, range from 1.2 times to as high as 30 times. Adopting log model solves the problem by linearizing the variation according to the order of the magnitude. Nonetheless, both linear and logarithmical values are reported by iHDMA for completeness. Further, the hematomas can be identified by setting a fixed threshold, for example, OD = 0.2. However, this could cripple the flexibility and extensions. Therefore we propose a probability based detection algorithm which will report the probability of hematoma on individual locations based on normalized reference scans and historical data. Considering the reference vector as the sample set, standard deviation is found. For a given probability of false alarm (pF A ) dynamic threshold (figure: 5.3) is set. The outliers from set of reference scans (the samples exceeding this dynamic threshold) are reported in terms of both OD and excess in times of standard deviation (zstd ).

55

Figure 5.3: In threshold detection, dynamic threshold is generated by maintaining constant probability of false alarm. A similar approach is taken and the Threshold Factor is in realtime controlled by the slider shown in the diagram.

Figure 5.4: Results after threshold detection.

56

Figure 5.5: Post detection integration is used to improve the accuracy, and automating detection. The key issue is to remove improbable scenarios similar to one shown in figure: 5.4). As it can be seen from the above image, the discrepancy is removed through post detection integration using algorithms such as BAM.

5.5

Post detection integration

The certitude of the presence of hematoma can be further enhanced by considering the neighboring samples. Because hematoma cannot exist as an isolated tiny dot, principle of locality could be applied to identify anomalies. The simplest of all is to select a set of n samples (typically n = 5) around the sample of interest (n − 2 to n + 2) and count the positive indications. If number of positive indications are greater than 50% or a given constant, the indication ppostdetectionintegration could be ascertained (Equation:2.2). This is coupled with the parameters derived earlier (either OD or standard deviation) to give a better estimate (Equation:5.5).

57

Ihematoma = ppostdetectionintegration × zstd

5.6

(5.5)

Application of Bidirectional Associative Memory (BAM)

While aforementioned probability based approach could yield satisfactory results it could be easily misled. For example, a vector of [1, 0, 1, 0, 1] could make the system ascertain a positive indication, while it is clearly recognizable that an alternate zeros and ones can mean some errors and not the presence of a hematoma. As introduced in section:2.5.3, BAM is an ANN model that can be used to detect patterns with machine learning. This could enable us to estimate the existence of hematoma with higher accuracy. It can be used to infer the presence of hematoma on a selected set of samples by learning from experience. Let, p1×n be the vector of selected samples and typically P1×n . We generate BAM vector Wn×n by, W =

R X

r r r T

q tp

(5.6)

r=1

Where, q is a constant found to minimize the energy of W and, t is the expected results generated from the past experience. Following the method, explained in section:2.5.3 presence of hematoma can be identified when a complete recall is available. However it must be noted BAM, being a semi-automatic algorithm using ANN, destroys the probability information prevailed thus far, and hence it would not be possible to generate the probability of existence further down.

58

5.7

Summary

We began the introduction by revisiting OD calculations. We also looked at the merits and demerits. Then we analyzed the characteristics of intensity values and established that continuous scan improves the outcome. We looked at few algorithms to normalize the reference count so that one-to-one mapping can be made between reference scan samples and active scan samples. We then introduced the application of threshold detection followed by post detection integration. Lastly, we discussed the use of an ANN model such as BAM for pattern matching. The next chapter presents the evaluation of the iHD system and discusses the results.

59

Chapter 6 RESULTS, EVALUATION AND DISCUSSION

In this chapter we discuss the clinical trial procedures, output of the individual components, performance metrics, and presentation of results.

6.1

6.1.1

Testing procedures and clinical trials

Trial Objectives and Purpose

The purpose of the trial is to assess the efficacy of the iHD and to obtain the statistical data of the area and concentration of haemoglobin in intracranial haemorrhages.

6.1.2

Trial Design

The patients undergoing surgery, due to an intracranial hematoma are intended to be used in the trial procedure. The trial procedure will happen between the events of taking the CT scan of the patient, and the starting of the surgery. To avoid bias, all the patients undergoing surgery due to intracranial hematoma within a specified time period (from 5pm 21st May to 8am 22nd May), will be taken for the trial.

60

The expected duration of subject participation is 5 minutes before undergoing the surgery and a follow up trial after the surgery is conducted with a duration of another 5 minutes, when possible. All patients undergoing the trial will be scanned with the iHD. There will be 2 areal scans done in each side of the head of the patient, and another 8 single scans on 8 specific areas of the head. The discontinuation criteria for this trial is, if the patient shows signs of external bleeding, or is the injured side of the patient cannot be touched without causing pain to the subject. The iHD project team will be accountable for conducting and managing the trial procedure.

6.1.3

Selection and Withdrawal of Subjects

• Subject inclusion criteria: The patient should be suspected of having an intracranial hematoma. The patient should be undergoing surgery. • Subject exclusion criteria: The patient having an external bleeding. The patient being touched on the injured side will result in affliction. • Subject withdrawal criteria: The patient having an external bleeding. The patient being touched on the injured side will result in affliction.

6.1.4

Treatment of Subjects

• The patient will be scanned in the injured side. • The patient will be scanned in the opposite side. • Singular scan values are obtained in 8 different locations in the head shown in figure:.3.2 61

6.1.5

Assessment of Efficacy

• Specification of the efficacy parameters: The following will be the main parameters that would be used to measure the efficacy of the device. – The difference in the optical density on the injured side and the reference side. – The correlation of the area of the hematoma from the CT scan and the detected region by the iHD. • Methods and timing for assessing, recording, and analyzing efficacy parameters: The efficacy parameters will be stored in electronic format, as an xml file, and will be analyzed later by the iHD project team.

6.1.6

Assessment of Safety

The safety regulations defined by FDA and discussed in section 2.4 have been followed in the design as illustrated in section 4.2.6, and will be abided by in conducting the tests as well.

6.1.7

Statistics

All the patients that satisfy the trial inclusion criteria at the Accident ward, National Hospital, Colombo during the time period mentioned will be used for the trial, and no sampling method will be used. The number of subjects planned to be enrolled is 20. This number is used considering the average rate of reported cases of intracranial hematomas in the National hospital and the statistical significance.

62

Data Handling and Recordkeeping : All the data of the trial will be kept and handled by the iHD project team, until they are published. Publication Policy : The data and records of the trial will be published by the iHD project team, at the completion of the trials.

6.2

6.2.1

Evaluation and Discussion

Sensor outputs

Light to voltage converter generates on average 700 mV when NIR light is sent through the healthy side of the skull at frontal area. This is comparable with the ambient noise of 50 mV . Therefore, the device is capable of detecting maximum of 1.15 in OD. On an unshaved head, the results have not been satisfactory and average results ranged between 100 mV and 150 mV . As the hair absorption is significant, it will be necessary to increase the power radiated. Therefore the device that is tuned for the use on a shaved head will be ineffective when there is hair absorption, in which case greater power is needed. ADC has a resolution of 10 bits with reference voltage of 5V. Hence average sensor output corresponds to a maximum of 143 of 1024 steps available, and therefore 8 least significant bits of 10 bit ADC output will be sufficient to detect the hematoma. Therefore on continuous scan only last 8 bits are stored for analysis.

6.2.2

Performance metrics

The complete process takes less than 15 minutes. Single scan takes on average 5 seconds and programmatically it is ended in 10 seconds. Further delays are 63

enforced to avoid accidental changes. Transfer of data from iHD to iHDMA over Bluetooth is almost instantaneous. Full set of calculations in iHDMA on a notebook with pentium AMD Turion 2.0 GHz processor and 2GB memory running Vista OS takes less than 2 seconds. The terminal consumes on average 700 mA during operation. A pair of lithiumion dual-cell can feed the device for more than 3 hours continuously. However the device will be put in standby mode and consumes lesser power normally. In practical situation, this translates to more than 30 scans per recharge.

6.3

Presentation of results

Firstly, acute differences in neighborhood and unequal sample count of reference scan samples are accounted for through normalization function.

Figure 6.1: Reference samples before normalization

Threshold detection is performed on the log ratio between scan and reference scan values. 64

Figure 6.2: Reference samples with normalization factor 4

Figure 6.3: The image illustrates the effect of normalization factor in smoothing reference scan values.

Final results are also visualized to provide a better interpretation.

65

Figure 6.4: The image depicts the effect on using different functions for normalization.

Figure 6.5: The image depicts the effect on using different functions for normalization.

66

Figure 6.6: Threshold detection is performed on the log ratio between scan and reference scan intensity values. The image shows the OD calculation.

Figure 6.7: The image demonstrates the use of BAM to filter out unlikely detections semiautomatically by matching patterns.

67

Figure 6.8: The image shows the visualization of the results on a subject with injury on left side of the head. Intensity variation on the injured side is shown by the red color component. Further positive indications that are confirmed by post detection integration - i.e. using BAM - are marked by white colored circles.

Figure 6.9: Quick scan is used to calculate the OD values on predefined locations for quick analysis by an expert. Under this mode, detection is made manually from OD values.

68

Chapter 7 CONCLUSION

The report discussed the concepts behind the design of intracranial hematoma detector, implementation, development of algorithm for semiautomatic detection, and evaluation. Intracranial hematoma detector is a non-invasive system to detect the presence of hemorrhages in brain with above reasonable accuracy semiautomatically. We summarize design challenges, contributions of the project and future directions for enhancement.

7.1

Challenges

We faced numerous challenges in designing the system, developing detection algorithms and evaluating the results.

• Power and safety: NIRS is a non-invasive detection technology used for brain imaging. However, the potential temperature rise due to heat transfer limits the maximum power that can be used. Although the primary transfer mode is conduction that can be prevented by proper isolation, radiated heat also plays a considerable role - 0.5o C at 50 mW/cm2 , that limits the power rating of the LEDs used. It must also be noted that solar radiation at NIR wavelength is of the same order, and hence contributes to significant noise. The dilemma is to increase received signal power, while keeping the source 69

power rating within the limits regulated by FDA. It is mainly achieved by pulsing the source current, and producing NIR light in burst of pulses which results in greater peak power while keeping the average power below the limit. • Compactness of the device: The size and design of the iHD should be portable as it is expected to be a handheld. The circuits are arranged in layers to stack them in smaller space. Also mobile phone batteries are used to provide enough power while saving space. • Detection and false alarm: The ability to detect a hematoma is inversely proportional to the probability of false alarm. Hence, the correct balance must be attained between both, and it is achieved by maintaining a variable factor, that is adjusted semiautomatically. • Field testings: The efficacy of the device can be only tested when it is deployed in real world scenario. The difficulty is in obtaining the consensus from the patients and relevant medical officers, and infrequency of arrival of patients with extracerebral hematomas over a short time span. The testings will be done.

7.2

Contributions

The report discussed the design and implementation of intracranial hematoma detector - iHD, and accompanying mobile application - iHDMA. Also it listed out the challenges faced in developing the solution. We itemize the contributions of the report below.

70

• Hardware device: iHD terminal is a simple microcontroller based system that uses bursts of NIR light energy as per the specifications that are partially configurable to measure the absorption differences. The system interconnects several subsystems including NIR transmitter and sensor, measures the intensity of NIR light at reception, stores the data and transfers to the mobile application. • Continuous scan:

In contrary to conventional methods of taking mea-

surements at particular points on the head and calculating OD, we adopted a novel way of using continuous scan from radar systems. It enables us to provide a visual imaging and a more descriptive picture of the details. • Semiautomatic detection:

To assist less experienced personnel, espe-

cially when it is used in field operations in the absence of experts, semiautomatic detection is developed. Here,threshold detection - another derivation from radar systems, is used with variable threshold factor. • BAM based filtering:

The detection might include improbable targets.

With machine learning, however, we can filter out them. BAM is an artificial neural network based tool that is used to detect patterns. BAM is adopted here for the aforementioned post detection identification.

7.3

Future directions

Although the system if fully functional, it could be enhanced in usability, accuracy and feature set. In our discussions with neurologists, we understood they desire better imaging features from this solution to substitute CT scans in peripheral hospitals. Currently the system scans along a line, i.e. linear in dimension. By implementing an array of LEDs and sensors it could be developed into a two-dimensional solution, which could generate a complete visual imaging of the brain. 71

The device can be made more compact, resembling a mobile phone device by using smaller surface mount components and advanced materials for chassis such as recyclable glass and aluminum. In order to keep the design simple and elegant it was decided to off-load all reporting functions to a mobile application. However, calculations such as OD value can be moved to iHD. Also to guarantee greater accuracy the signal can be modulated by a predefined frequency and at the reception a matched filter can be used to extract the signal from ambient noise. It might result in more accurate detection in the presence of significant ambient noise such as direct solar radiation.

72

BIBLIOGRAPHY

[1] L.A. Al-Gwaiz and H.H. Babay. The diagnostic value of absolute neutrophil count, band count and morphologic changes of neutrophils in predicting bacterial infections. Medical Principles and Practice, 16(5):344–347, 2007. ISSN 1011-7571. URL http://www.karger.com/DOI/10.1159/000104806. [2] Yoshinori Iwatani. The medical technologist as a key professional in medical care in the 21st century. Rinsho Byori, 56(10):915–923, Oct 2008. [3] Esther L Yuh, Alisa D Gean, Geoffrey T Manley, Andrew L Callen, and Max Wintermark. Computer-aided assessment of head computed tomography (ct) studies in patients with suspected traumatic brain injury. J Neurotrauma, 25(10):1163–1172, Oct 2008. doi: 10.1089/neu.2008.0590. URL http://dx.doi.org/10.1089/neu.2008.0590. [4] L. L. Teunissen, G. J. Rinkel, A. Algra, and J. van Gijn. Risk factors for subarachnoid hemorrhage: a systematic review. Stroke, 27(3):544–549, Mar 1996. [5] Deaths gions,

by

cause,

estimates

changing

history.,

for

sex 2002. World

and

mortality The

Health

world

stratum

in

health

report

Organization,

2004.

who

re-

2004

-

URL

http://www.who.int/entity/whr/2004/annex/topic/en/annex_2_en.pdf. [6] Birsen Yazc Il-Young Son. Near infrared imaging and spectroscopy for brain activity monitoring.

73

[7] S. P. Gopinath C. S. Robertson and B. Chance. Use of near infrared spectroscopy to identify traumatic intracranial hematomas. J. Biomed. Opt, 2: 3141, 1997. [8] D. Kushner. Mild traumatic brain injury: toward understanding manifestations and treatment. Arch Intern Med, 158(15):1617–1624, 1998. [9] A. Mishra and S. Mohanty. Contre-coup extradural haematoma : a short report. Neurol India, 49(1):94–95, Mar 2001. [10] Sanders MJ and McKenna K. Head and Facial Trauma. Mosby, 2nd revised edition, 2001. Chapter 22. [11] Traumatic physiology:

brain

injury

eMedicine

(TBI)

-

physical

definition, medicine

epidemiology, and

patho-

rehabilitation.

http://emedicine.medscape.com/article/326510-overview.

URL

http://emedicine.medscape.com/article/326510-overview. [12] A. Villringer and B. Chance. Non-invasive optical spectroscopy and imaging of human brain function. Trends Neurosci, 20(10):435–442, Oct 1997. [13] A. Sauaia, F. A. Moore, E. E. Moore, K. S. Moser, R. Brennan, R. A. Read, and P. T. Pons. Epidemiology of trauma deaths: a reassessment. J Trauma, 38(2):185–193, Feb 1995. [14] Eugene Park, Joshua D Bell, and Andrew J Baker. Traumatic brain injury: can the consequences be stopped? CMAJ, 178(9):1163–1170, Apr 2008. doi: 10.1503/cmaj.080282. URL http://dx.doi.org/10.1503/cmaj.080282. [15] P. L. Hope and J. Moorcraft. Magnetic resonance spectroscopy. Clin Perinatol, 18(3):535–548, Sep 1991. [16] Patrick J. Treado, Ira W. Levin, and E. Neil Lewis. Near-infrared acoustooptic filtered spectroscopic microscopy: A solid-state approach to chemical 74

imaging. Applied Spectroscopy, 46:553–559(7), 1 April 1992. doi: doi:10. 1366/0003702924125032. URL http://www.ingentaconnect.com. [17] S. P. Gopinath, C. S. Robertson, R. G. Grossman, and B. Chance. Nearinfrared spectroscopic localization of intracranial hematomas. J Neurosurg, 79(1):43–47, Jul 1993. [18] CLAUDIA

S.

ROBERTSON,

BRITTON CHANCE.

SHANKAR

P.

GOPINATH,

and

A new application for near-infrared spec-

troscopy:

Detection of delayed intracranial hematomas after head

injury.

Journal

of

Neurotrauma,

12(4):591–600,

1995.

URL

http://www.liebertonline.com/doi/abs/10.1089/neu.1995.12.591. [19] Y. Ito, R. P. Kennan, E. Watanabe, and H. Koizumi.

Assessment of

heating effects in skin during continuous wave near infrared spectroscopy. J Biomed Opt, 5(4):383–390, Oct 2000.

doi: 10.1117/1.1287730.

URL

http://dx.doi.org/10.1117/1.1287730. [20] Alper Bozkurt and Banu Onaral.

Safety assessment of near in-

frared light emitting diodes for diffuse optical measurements. Eng Online, 3(1):9, Mar 2004.

doi:

10.1186/1475-925X-3-9.

Biomed URL

http://dx.doi.org/10.1186/1475-925X-3-9. [21] K. G. Murphy, J. A. Secunda, and M. A. Rockoff. Severe burns from a pulse oximeter. Anesthesiology, 73(2):350–352, Aug 1990. [22] D. B. Sobel. Burning of a neonate due to a pulse oximeter: arterial saturation monitoring. Pediatrics, 89(1):154–155, Jan 1992. [23] Guyton AC. Textbook of Medical Physiology. Saunders, 2000. [24] Y. F. Wang, Jr. Cruz, J. B., and Jr. Mulligan, J. H. Two coding strategies for bidirectional associative memory. 1(1):81–92, March 1990. doi: 10.1109/ 72.80207. 75

[25] B. Kosko. Bidirectional associative memories. 18(1):49–60, Jan.–Feb. 1988. doi: 10.1109/21.87054. [26] Y. F. Wang, Jr. Cruz, J. B., and Jr. Mulligan, J. H. Guaranteed recall of all training pairs for bidirectional associative memory. 2(6):559–567, Nov. 1991. doi: 10.1109/72.97933. [27] ”Texas light

Instruments”. to

voltage

Optoelectronic -

opt101.

products

Datasheet.

URL

http://focus.ti.com/docs/prod/folders/print/opt101.html. [28] Q. Zhang, H. Ma, S. Nioka, and B. Chance. Study of near infrared technology for intracranial hematoma detection. J Biomed Opt, 5(2):206–213, Apr 2000. doi: 10.1117/1.429988. URL http://dx.doi.org/10.1117/1.429988.

76

Appendix A IHD SCHEMATIC

Figure A.1: Schematic diagram of iHD terminal. Signal generation and sensor subsystem and Communication subsystem are shown in the diagram.

77

Figure A.2: Schematic diagram of iHD. Microcontroller and power management subsystem are show in the diagram

78

Appendix B COMMUNICATION PROTOCOL SPECIFICATION

Version : 0.1 The document describes the protocol that is adopted to be used for the communication between the remote intracranial hematoma detector (iHD) and Medical Application (iHDMA) running from a Personal Computer, Notebook or Mobile Computer. Please note “iHD” and “iHD terminals” are used interchangeably to denote the device. The following media are currently supported without modifications • Bluetooth • Serialport In addition, other media may be supported, but have neither been tested nor documented.

B.1

Introduction

The system is planned to accommodate multiple devices non-concurrently, and hence uses Host-Slave architecture. The iHDMA will always be a host and iHD will always be a slave. Therefore, only iHDMA can initiate communication by querying the iHD, and the iHD is obliged to respond. 79

The atomic unit of the data is a word (16 bits) which is in big-endian format, i.e. the higher order byte is sent first. All the messages are encapsulated between a message-starting header and messageending header. Further frame headers wrap all data frames. One cycle of communication includes the command initiated by iHDMA and the reply from iHD. Commands can be one of the following • Request Settings • Update Settings • Request Data • Ping

B.1.1

Structure

As aforementioned, a command or reply message is wrapped by headers. The body can be of any number words long. Typical command or reply message

Figure B.1: The header identifies the type of the word (i.e. a Message Header) and type of the command (Ping, Request Data, Request Settings or Update settings), and the body is supplemented with any parameters if available. Hence, the body may be absent when it is neither required by the command nor available. The command or reply is ended with message-ending header which is 0xFFFF for all messages.

80

B.2

Message-starting Header

A 16 bits long word of the following format will serve as a message-starting header :

Figure B.2: A 16 bit word serving as a message-starting header.

There can be four types of commands :

Figure B.3: Four types of words used for message starting header of commands.

And the corresponding replies are as follows :

Figure B.4: Four types of words used for message starting header of replies.

B.3

Message Body

For a command, message body is used to pass the relevant parameters. For a reply, message body consists of information that is queried through the command and can vary depending on the type of the command. All possible message body combinations are explained below. 81

Figure B.5: Message body of ping reply. Ping reply can be anything, no restriction is imposed on its format, but still it has to be divisible by 3.

Figure B.6: Message Body of update settings and request settings reply. Replies are the same and both return all setting parameters, one per word (Settings reply can be anything, no constraints on its format, but it cannot be 0xFFFF at any time to distinguish from Message-ending Header).

Figure B.7: Message body of request data reply. Data is packetized one level further into frames. A typical request data - reply will be of this format, hence making the frame a packet.

B.4

Message-ending Header

Figure B.8: Message-ending Header. All 16 bits are set to 1s and thus easily identified. Hence, data packets of any type, settings-replies, frame headers and ping-replies cannot have all 16 bits set.

B.5

Frame

A frame can be either a frame-idle or data-frame

Figure B.9: Frame-idle.

82

Figure B.10: Data-Frame consists of a Frame-starting header, data packets, and a frame-ending header.

B.6

Example Word Patterns

Figure B.11: Typical word. The first 4 bits of a word are used to identify the type of the word. This is possible because core information is of maximum 10 bits long (in the case of ADC output) and thus avoids unnecessary escaping and processing that will be otherwise needed.

B.6.1

Word Types

Determined by the first 4 bits of any word. 1. Message header (1111) 2. Frame Header (1011) 3. Frame idle (1101) 4. Data Packet (0000) In addition, special word - Message-ending header (0xFFFF) will be determined by the whole word (all 16 bits)

83

B.6.2

Message ID

Determined by the last 4 bits of the message-starting headers. 1. Ping (0001) 2. Request Data (0010) 3. Request Settings (0100) 4. Update Settings (1000)

B.6.3

Command or Reply

Determined by the 3rd bit of the second word in a message-starting header. 1. Command (0) 2. Reply (1)

B.6.4

Frame ID

Determined by the last 4 bits of the frame-starting header. 1. Real time scan (0001) 2. Real time calibration (0010) 3. Stored OD Values (a) ODa (0100) (b) ODb (1000) (c) ODc (1100) (d) ODd (1110) 84

B.7

Summary

This protocol is designed with the scenario of multiple iHDs communicating with iHDMA non-concurrently but simultaneously. Further, a typical communication pattern of Pinging device, Request settings, Update settings, and Data acquisition is assumed. The packets are formed from the atomic unit of word (of 16 bits with first 4 bits identifying the type of the word).

85

Appendix C DATASHEETS

Figure C.1: Datasheet of PIC 18F452 - the microcontroller used in iHD terminal.

86

Figure C.2: Pin diagrams of PIC 18F452 - the microcontroller used in iHD terminal.

87

Figure C.3: Specifications of OPT101 - the light to voltage converter used in iHD terminal.

88

Figure C.4: OPAMP characteristics of OPT101 - the light to voltage converter used in iHD terminal.

89

Figure C.5: Datasheet of 760/850 nm narrow bandwidth high power NIR LEDS used in iHD terminal..

90

Appendix D BIDIRECTIONAL ASSOCIATIVE MEMORY (BAM)

We will first look at the derivation and the algorithm of BAM. Then we will present the conditions for complete recall. Given input data vector r Pn , target data vector r tn , where r denotes the rth sample, and n denotes the nth element, first step is to generate BAM vector Wn×n

W=

R X

r r r

q t pT

(D.1)

r=1

It is followed by the following recurrent steps for a new input p, r (0)

c

r (k)

= f (Wr c(k) )

a

r (k+1)

c

= rp

= f (WT r a(k) )

(D.2)

and finally when equilibrium is achieved, r (N )

= rp

r (N )

= rt

c

a

91

(D.3)

that happens when energy E is minimal. E(r P, r t) = −r aT Wr c

(D.4)

The equilibrium is ascertained when vectors c and a remain unchanged over iterations. Although Kosko’s BAM (r q = 1) would result in equilibrium often, Yeou-Fang Wang et al [26] showed that complete recall is only possible when E is in its local minimum. For p and p1 are different by one Hamming Distance, a pair (l p, l t) can be recalled iff nl = l tT r tr pT l p − l t1T r tr pT l p1 ∀l

R X

r

qnl ≥ 0

r=1

92

(D.5) (D.6)