Cole Gray

The Effects of Gray Scale Image Processing on Digital Mammography Interpretation Performance1 Elodia B. Cole, MS, Etta D. Pisano, MD, Donglin Zeng, PhD, Keith Muller, PhD, Stephen R. Aylward, PhD, Sungwook Park, Cherie Kuzmiak, DO, Marcia Koomen, MD, Dag Pavic, MD, Ruth Walsh, MD, Jay Baker, MD, Edgardo I. Gimenez, MD, Rita Freimanis, MD

Rationale and Objectives. To determine the effects of three image-processing algorithms on diagnostic accuracy of digital mammography in comparison with conventional screen-film mammography. Materials and Methods. A total of 201 cases consisting of nonprocessed soft copy versions of the digital mammograms acquired from GE, Fischer, and Trex digital mammography systems (1997–1999) and conventional screen-film mammograms of the same patients were interpreted by nine radiologists. The raw digital data were processed with each of three different image-processing algorithms creating three presentations—manufacturer’s default (applied and laser printed to film by each of the manufacturers), MUSICA, and PLAHE—were presented in soft copy display. There were three radiologists per presentation. Results. Area under the receiver operating characteristic curve for GE digital mass cases was worse than screen-film for all digital presentations. The area under the receiver operating characteristic for Trex digital mass cases was better, but only with images processed with the manufacturer’s default algorithm. Sensitivity for GE digital mass cases was worse than screen film for all digital presentations. Specificity for Fischer digital calcifications cases was worse than screen film for images processed in default and PLAHE algorithms. Specificity for Trex digital calcifications cases was worse than screen film for images processed with MUSICA. Conclusion. Specific image-processing algorithms may be necessary for optimal presentation for interpretation based on machine and lesion type. Key Words. Digital mammography; ROC curve; image processing. ©

AUR, 2005

The separation of acquisition from display, in full-field digital mammography, allows for optimization of image

Acad Radiol 2005; 12:585–595 1 From the Department of Radiology and Lineberger Comprehensive Cancer Center (E.B.C., E.D.P.) and Department of Radiology (S.R.A.), University of North Carolina, CB#7515, Radiology Research Labs, 106 Mason Farm Road, Chapel Hill, NC 27599; University of North Carolina, Department of Biostatistics, Chapel Hill, NC (D.Z., K.M.); WiBro Terminals Labs, Telecommunication R&D Center, Samsung Electronics Co., Gyenoggi-do, Korea (S.P.); Department of Radiology, University of North Carolina, Chapel Hill, NC (C.K., M.K., D.P.); Department of Radiology, Duke University Medical Center, Durham, NC (R.W., J.B., E.I.G.); Department of Radiology, Wake Forest University, Winston-Salem, NC (R.F.). Received October 28, 2004; revision received and accepted January 6, 2005. Supported by Susan G. Komen Foundation Grant. Address correspondence to: E.B.C. e-mail: [email protected]

© AUR, 2005 doi:10.1016/j.acra.2005.01.017

display at each point in the image formation chain. Before display of the final image, some type of image processing is applied to the raw digital mammographic image. Ideally, the application of an image processing algorithm will function to improve the visibility of lesions rather than just improving the aesthetic appeal of images. Several studies have specifically evaluated the effect of various types of image processing algorithms on radiologist performance (1–7). Image processing can take place in several places in the imaging chain: at the acquisition station, at the review station, or in between. Generally, the end user presses a predefined button that will display a processed image. Beyond the individuals who actually developed the algorithms or were responsible for implementing them in the various digital mammography systems, very little is 585

COLE ET AL

Academic Radiology, Vol 12, No 5, May 2005

Table 1 Lesion Type and Cancer Status for All 201 Cases Calcifications

Fischer GE Trex

Masses

Benign

Malignant

Benign

Malignant

Normals

16 10 22

8 9 8

22 14 20

14 10 13

14 9 12

known about why specific algorithms were chosen or under what conditions they are applied (ie, Is the specific image processing algorithm more suitable for mass characterization or calcification characterization? Is the specific image processing algorithm better suited for application to specific breast densities? Does the specific algorithm improve lesion visibility or radiologist diagnostic accuracy?). The purpose of this study was to quantify the effect of algorithms applied to digital mammograms of women with dense breasts acquired from three different digital mammography systems on radiologist performance when compared with screen-film mammography.

METHODS AND MATERIALS A total of 201 cases were obtained from the International Digital Mammography Development Group image archive. This case set is described in detail in a previous article (8). Table 1 shows the breakdown of the cases by lesion type and cancer status. Each case consisted of four standard view digital and screen-film mammograms of the same patients. The patients were enrolled and imaged at one of seven institutions under the protocols of a previous study (8). Three digital mammography systems were tested: the SenoScan (Fischer Imaging Corp, Denver, CO), the Senographe 2000D (General Electric Corp., Waukesha, WI), and the Trex Digital Mammography System (Hologic Inc., Bedford, MA). Each patient was imaged on only one type of digital mammography system. Available screen-film mammography systems at each site were used to obtain the screen-film mammograms during this study. Three algorithms were applied to each of the digital images: MultiScale Image Contrast Amplification (MUSICA), a commercially available algorithm (AgfaGevaert N.V., Belgium); Power Law Adaptive Histogram Equalization (PLAHE), an algorithm developed at our institution; and the manufacturer’s recommended or default algorithm, the algorithms recommended and implemented by each manufacturer at the time this data was 586

collected between 1997 and 1999. The Food and Drug Administration had not yet clinically approved any of the full-field digital mammography systems at the time these data were collected. Each was in various stages of the clinical trials process. However, the machines used were the same as the ones ultimately approved for clinical use by the Food and Drug Administration. The MUSICA and PLAHE algorithms were applied to all 201 cases. Each manufacturer’s default algorithm was applied to the cases acquired using that manufacturer’s digital mammography system. The algorithms were not always applied successfully because of incorrect interpretation of image parameters by the image processing software. Image format standardization had not yet been implemented at the start of this study. Therefore, across platforms, the image formats varied considerably from one manufacturer to the next and even within manufacturer. Cases acquired from one prototype system differed in format (byte order, image size, gray-scale representation) from another provided by the same manufacturer. Table 2 shows the distribution of cases that were successfully processed and thus used for the final reader study. The application of an algorithm was considered successful for an image if the resultant image was interpretable in the judgment of a board-certified Mammography Quality Standards Act (MQSA)-qualified breast imaging radiologist (E.D.P.). The processing success rate was 93.5% (188/201) for the MUSICA algorithm, 89.1% (179/201) for the manufacturer’s default algorithms, and 67.2% (135/201) for the PLAHE algorithm. The default algorithms were applied by each manufacturer to the images acquired on their respective machines by applications specialists. The default processed digital mammograms were subsequently printed to film for radiologist interpretation in the reader study. Agfa Corporation provided a stand-alone version (version 2.0.0) of their MUSICA software for use in this study. Research personnel at our institution applied PLAHE and MUSICA algorithms to all digital mammograms. For radiologist interpretation in the reader study, PLAHE and MUSICA cases were displayed in soft copy format using a mammography soft copy review workstation (Sun platform, two high-resolution high contrast monitors) running Mammoview software, as described elsewhere (9). Nine readers participated in the reader study. All had mammography experience, with an average experience of 11 years (range 1–18 years). Eight of nine readers were American Board of Radiology (ABR) -certified attending radiologists. One reader was a breast imaging fellow. The

IMAGE PROCESSING AND DIGITAL MAMMOGRAPHY

Academic Radiology, Vol 12, No 5, May 2005

Table 2 Lesion Type and Cancer Status Case Distributions for Successfully Processed Cases Used in Reader Study per Image Processing Algorithm

Benign Fischer

Malignant

Calcifications Mass Calcifications Mass

Normal Fischer totals

Benign GE

Malignant

Calcifications Mass Calcifications Mass

Normal GE totals

Benign Trex

Malignant

Calcifications Mass Calcifications Mass

Normal Trex totals

MUSICA

PLAHE

Default

16 20 8 13 10 67

16 19 8 13 10 66

16 20 8 13 8 65

10 13 9 8 9 49

8 3 3 5 4 23

10 12 9 9 8 48

22 19 8 12 11 72

14 10 4 10 8 46

22 15 7 12 10 66

MUSICA: MultiScale Image Contrast Amplification; PLAHE: Power Law Adaptive Histogram Equalization.

Figure 1. Time spent reading mammograms is between 10 and 40 hours for our nine readers. On average, 55% of their time is spent on screening mammography and 45% is spent on diagnostic mammography.

average amount of time spent by the readers as a group interpreting mammograms per week was 28.11 hours (range 10 – 40 hours/week; 55% screening, 45% diagnostic) (Fig 1). Three readers were assigned to read all readable cases for MUSICA and to the corresponding screen film mam-

mograms. Another three readers were assigned to read all readable cases for PLAHE and to the corresponding screen film mammograms. Yet another three readers were assigned to read all readable cases for default and to the corresponding screen film mammograms. All nine readers began the reader study by completing the screen film mammograms interpretations first on a standard radiology multiviewer appropriately masked for mammography. After a minimum 4-week washout period, each of the nine readers then read the digital mammograms to which they were assigned. The readers were provided with structured paper forms to facilitate consistent reporting of mammographic findings. The readers reported clinically significant lesion location (breast, o’clock location, anteroposterior depth), lesion-specific Breast Imaging Reporting and Data Systems (BI-RADS) (American College of Radiology, Reston, VA) characteristics, and a probability of malignancy based on a 5-point scale (1 definitely not malignant, 2 probably not malignant, 3 possibly malignant, 4 probably malignant, and 5 definitely malignant). Responses for lesion location and probability of malignancy were compared with ground truth based on biopsy or 1-year follow-up using methods described elsewhere (8). If the 587

COLE ET AL

reader determined there were no significant findings, no probability of malignancy was recorded. Three primary outcomes were analyzed separately: 1) area under the receiver operating characteristic (ROC) curve (AUC) from a nonparametric ROC analysis of cancer/no cancer; 2) sensitivity, with no findings and 1–2 defined as “benign” and 3–5 as “malignant”; and 3) specificity, with no findings and 1–2 defined as “benign” and 3–5 as “malignant.” For each outcome, a general linear multivariate model analysis was used to fit data and repeated measures tests based on the Geisser-Greenhouse test were used to test the significance of main effects as well as interactions among the factors of machines, lesion types, and digital display methods. In the multivariate linear model, the outcome was the repeated measurements of outcomes (AUC difference, sensitivity of specificity differences) of a particular reader in different conditions of three machines and two lesion types; the design matrix included a three-level categorical predictor of three display methods and the baseline performance with film screen was also adjusted in the model. “Main effects” refer to the overall effects of each factor. The interaction factors were put in the model at the same time. All the results from the residual analysis indicated a goodness of fit and showed evidence of validity of normality assumption. Thus the identity link function we used in the model fits the data well. Data analysis was performed using SAS Software, Version 8.0 (SAS Institute, Cary, NC).

RESULTS ROC curve analysis was performed. The outcomes measured were area under the ROC curve, sensitivity, and specificity. Figure 2 shows the ROC curves for digital and screen film mammography by machine type and lesion type across all image processing algorithms. The ROC curves for digital and screen film mammography by machine type and lesion type per image processing algorithm are displayed in Fig 3–5.

AUC The primary outcome of interest was the area under the ROC curve difference between digital mammography and screen film mammography. Table 3 reports the average of this difference by machine type and lesion type. The AUC for digital mammography was worse than screen film for all machine and lesion combinations ex588

Academic Radiology, Vol 12, No 5, May 2005

cept for the Trex mass cases. The variance of the AUC difference of digital and screen film mammography was relatively large between Fischer digital mammograms and screen film mammograms. Likewise, the variance of the AUC difference was relatively large between GE digital mammograms and the screen-film mammograms (Fig 6). Tests based on the multivariate analysis of AUC data show that, on average, the digital mammograms produced by the GE system showed a significantly worse AUC than the screen film mammograms (P ⫽ .009). In addition, the digital mammograms across all machine types showed a worse AUC for both mass (P ⫽ .007) and calcifications (P ⫽ .042) lesion types (Table 3). To study which of the machine type, lesion type, and processing method variables produce the difference between digital mammography and screen film mammography, multivariate analysis was conducted of AUC data including the interactions among these factors. Trex digital mammography produced a larger AUC difference with screen film mammography than Fischer (P ⫽ .019), but no significant difference was observed between GE and Fischer (P ⫽ .193). There was no significant difference seen among the three processing methods (MUSICA, PLAHE, default) (P ⫽ .0.168, P ⫽ .485) or between the two lesion types (P ⫽ .0.822). The resulting 95% confidence intervals for AUC difference are shown in Fig 6. GE mass cases processed with each of the three image processing methods produced a significantly worse AUC than screen film. Trex mass cases processed with default produced significantly better AUC than screen film. All other digital combinations were statistically indistinguishable from screen film. Sensitivity The sensitivity difference between digital mammography and screen-film mammography was calculated. Table 3 reports the average of this difference by machine type and lesion type and shows that the sensitivity from GE digital mammography was lower than the sensitivity from screen film mammography. However, the opposite was seen when digital mammograms were generated from the Trex system for both lesion types and for the Fischer system for calcifications. The variation of the sensitivity difference was relatively large for the digital mammograms generated by Fischer and GE systems and screen film mammograms. Tests based on multivariate analysis of sensitivity data shows that, on average, GE digital mammography had significantly less sensitivity than screen film mammography (P ⫽ .011); and, on average, the digital mammo-

Academic Radiology, Vol 12, No 5, May 2005

IMAGE PROCESSING AND DIGITAL MAMMOGRAPHY

Figure 2. Area under the receiver operating characteristic curves by machine and lesion type across all image processing algorithms; value presented in parentheses for each modality.

grams across all machine types had less sensitivity for mass lesion type (P ⫽ .014) (Table 1). To study which combinations of the machine type, lesion type, and image processing algorithm factors produce the difference between digital mammography and screen film mammography, multivariate analysis was conducted of sensitivity data including the interactions among these factors. Trex produced a larger variation in sensitivity difference with screen film mammography than Fischer digital (P ⫽ .005), but no significant difference was

observed between GE and Fischer. There was no significant difference seen between the three processing methods (P ⫽ .837 and P ⫽ .745) and between the two lesion types (P ⫽ .229). The resulting 95% confidence intervals for sensitivity difference are shown in Fig 7. GE mass cases processed with each of the image processing algorithms had significantly lower sensitivity than screen film. There was no statistically significant sensitivity difference between all other remaining digital combinations and screen film. 589

COLE ET AL

Academic Radiology, Vol 12, No 5, May 2005

Figure 3. Area under the receiver operating characteristic curves by machine and lesion type averaged across the three default readers; value presented in parentheses for each modality.

Specificity The specificity difference between digital mammography and screen film mammography was calculated. Table 3 reports the average of this difference by machine type and lesion type. It shows that the specificity from digital mammography was less than the specificity from screen film mammography across all machine types for calcifications. Tests based on multivariate analysis of specificity data show that on average, the digital mammograms produced by the Fischer digital system gave significantly less specificity than screen film mammograms (P ⫽ .044); and, on 590

average, the digital mammograms across all digital systems gave less specificity for calcifications (P ⫽ .003) (Fig 8). To study which combinations among the machine type, lesion type, and processing method factors produced the difference between digital mammography and screen film mammography, multivariate analysis was conducted of specificity data including the interactions among these factors. The mass cases produced larger variation in specificity difference between digital mammography and screen film mammography than the calcifications cases (P ⫽ .002), but no significant dif-

Academic Radiology, Vol 12, No 5, May 2005

IMAGE PROCESSING AND DIGITAL MAMMOGRAPHY

Figure 4. Area under the receiver operating characteristic curves by machine and lesion type averaged across the three MultiScale Image Contrast Amplification readers; value presented in parentheses for each modality.

ference was observed among the three machine types. No difference was seen among three processing methods and between the two lesion types. The resulting 95% confidence intervals for specificity difference are shown in Fig 8. The Fischer calcifications cases processed with default and PLAHE, and Trex calcifications cases processed with MUSICA produced significantly lower specificity than did screen film. All remaining digital combinations showed no difference in specificity from screen film.

DISCUSSION Radiologist sensitivity and specificity are dependent not only on the interpretation skill of the radiologist, but also to a certain extent on just how visible lesions actually are. First, the lesion must be distinguishable from the surrounding background (normal breast tissue) to be detected mammographically. Similarly, diagnosis of a lesion is only made when the lesion’s features can be classified. Image-processing algorithms are applied to digital mammograms to alter 591

COLE ET AL

Academic Radiology, Vol 12, No 5, May 2005

Figure 5. Area under the receiver operating characteristic curves by machine and lesion type averaged across the three Power Law Adaptive Histogram Equalization readers; value presented in parentheses for each modality.

the visual presentation of the digital mammograms through manipulation of the gray-scale values of the pixels making up the mammographic image. Ideally, this manipulation, be it point (individual pixel), area (clustered groups of pixels), or global (whole image) would lead to an image where lesions are more distinguishable from normal tissue to allow the radiologists to use their interpretation skill. The lesion has to be seen before it can be interpreted. An image-processing algorithm’s effect on an input image will be dependent on the spatial resolution of the input image, the contrast resolution available in the input image, the overall breast 592

density of the patient, and the lesion type if there are findings. There are several studies that have addressed the effects of spatial resolution on calcification detection in digital mammography (10 –12) for 100-micron digitized images. None showed improved sensitivity for digital mammography over screen film. With increased spatial resolution, microcalcifications should be better visualized. However, this resolution does not alone provide the ability to see smaller features such as calcifications; the available contrast resolution is also important. We would expect calcifications to be better visu-

IMAGE PROCESSING AND DIGITAL MAMMOGRAPHY

Academic Radiology, Vol 12, No 5, May 2005

Table 3 Main Effects by Machine Type and Lesion Type of AUC difference, Sensitivity Difference, and Specificity Difference Between Digital and Screen-Film Mammography AUC difference

Sensitivity difference

Specificity difference

Factor

Type

Mean (SD)

P value

Mean (SD)

P value

Mean (SD)

P value

Machine

Fischer GE Trex Calc Mass

⫺0.042 (0.037) ⫺0.205 (0.043) 0.026 (0.012) ⫺0.077 (0.026) ⫺0.070 (0.014)

.321 .009* .091 .042* .007*

⫺0.029 (0.048) ⫺0.255 (0.056) 0.082 (0.032) ⫺0.032 (0.044) ⫺0.102 (0.024)

.581 .011* .061 .508 .014*

⫺0.061 (0.023) ⫺0.037 (0.039) ⫺0.073 (0.038) ⫺0.129 (0.024) 0.015 (0.040)

.044* .398 .113 .003* .725

Lesion

AUC: area under the receiver operating characteristic curve. *Indicates statistically significant difference between digital and screen film mammography.

Figure 6. 95% confidence intervals for area under the receiver operating characteristic difference between digital and screen film mammography. Each vertical interval represents the machine type, image processing type, and lesion type combination seen by three readers. The intervals marked with asterisks were statistically significant.

alized on screen film compared with digital mammography based on the higher spatial resolution for screen film, but the increased contrast difference inherent with digital mammography is thought to compensate (13). Based on that assumption, we would also expect the sensitivity difference between digital and screen film mammography to decrease as the pixel size of the digital image gets smaller and smaller (increasing spatial resolution). These trends were found in our study, in which calcification visualization on these digital mammograms was worse than screen film overall, with a narrowing of the difference between digital mammography and screen film mammography for sensitivity with increasing spatial resolution. The Trex digital system, which had 40 microns per pixel size (highest spatial resolution system tested) and contrast resolution of 14 bits (16,384 distinct gray levels), had the

Figure 7. 95% confidence intervals for sensitivity difference between digital mammography and screen film mammography. Each vertical interval represents the machine type, image processing type, and lesion type combination seen by three readers. The intervals marked with asterisks were statistically significant.

best calcification sensitivity performance relative to screen film (Fig 7). The Fischer system was next with its 50 microns per pixel size and contrast resolution of 12 bits (4,096 distinct gray levels), and GE was third with spatial resolution of 100 microns per pixel size (lowest spatial resolution system tested) and contrast resolution of 15 (32,768 distinct gray levels) for one GE prototype and 16 bits (65,536 distinct gray levels) for another. An interesting finding in our study was calcifications sensitivity performance in certain instances was slightly better than the mass sensitivity performance, for which we believe lesion subtlety (the visibility of the lesion in relation to the surrounding breast tissue) was a factor. The overall breast density for each patient whose images were included in this study was at least heterogeneously dense (8), making this a difficult case set to interpret. The overall breast density has been shown to affect radiologist 593

COLE ET AL

Figure 8. 95% confidence interval for specificity difference between digital mammography and screen film mammography. Each vertical interval represents the machine type, image processing type, and lesion type combination seen by three readers. The intervals marked with asterisks were statistically significant.

ROC performance, showing decreased sensitivity and specificity (14) with increased breast density. This effect is true of screen film mammography and this would be true of digital mammography as well, unless significant visualization were possible through the larger contrast range. But although the maximum range of contrast values is set per machine type, the actual breast tissue composition and dose dictates the range of contrast values for each individual patient. The relative difference between each digital mammography system and screen film mammography for specificity is about the same. Specificity difference between all the digital machine types and screen film do not vary much with each of the image processing algorithms applied, regardless of lesion type. Both PLAHE and manufacturer’s default seem to be specifically optimized for improved characterization of masses across each of the machine types. A small relative difference was noticed in specificity, regardless of machine type for calcifications (Fig 8). This difference would be expected, given that the images used in this study were screening mammograms and not diagnostic magnification or spot compression views, in which better characterization of calcifications is possible. In assessment of the specificity (the ability to distinguish lesion features that are suspicious for cancer from lesion features not suspicious for cancer), it was expected that digital mammography would perform better for specificity by benefit of the digital mammograms being acquired after the lesion was found with screen film mammography. There was a smaller mean difference between digital mammography and screen film and there was less variability among readers in regard to specificity difference between digital 594

Academic Radiology, Vol 12, No 5, May 2005

mammography and screen film mammography. This specificity difference improvement is probably due to the fact that the digital mammograms for this study were obtained with the location of the lesion known by the technologist and presumably optimally positioned and compressed to show the characteristics of the lesion more clearly. Digital mammography in general is considered more specific than screen film mammography based on previous studies (8,15,16). However, none of those studies, or the study reported here, looked at randomly selecting the order of the imaging modalities. However, there is a study comparing digital mammography to screen film mammography that did include randomization of the order of imaging modalities— the Digital Mammography Imaging Screening Trial (DMIST), which was open to accrual between 2001 and 2003 and is currently in analysis—should resolve whether the increased specificity seen with digital is simply temporal. Limitations Although there were 201 cases included in this study, image-processing failures resulted in a lower number of cases being usable for the reader study. We suspect that the use of more stable digital mammography systems now commercially available and the widespread use of image format standards in medical imaging, such as Digital Imaging and Communications in Medicine (10), which was not available for all machines at the time of this study, would have led to better input image consistency in this study and fewer failures of the algorithms, especially PLAHE. More than 96% (52/54) of the PLAHE-processed cases were not available because of image-formatting variations within machine types that were not anticipated at the time of algorithm development. The three image-processing algorithms tested in this study— default, PLAHE, and MUSICA— had all performed well in previous studies where they were compared with other image-processing algorithms (3,8), although not necessarily better than screen film. There certainly could be better algorithms, even improved versions of the ones tested here, that could lead to different results. The robustness of algorithms and the parameter settings used can also be the cause of image display differences. All of the algorithms tested had some flexibility in regard to optimizing parameter settings. The possibilities in some instances were limitless. We conducted a small preference study by one expert radiologist to determine the optimal parameter settings for the MUSICA and PLAHE algorithms to be used for the cases included in this study. These optimal parameter settings were selected based on the radiologist’s impression of the quality of the image data in side-by-

Academic Radiology, Vol 12, No 5, May 2005

side comparison to screen film. The most consistently preferred parameters across both lesion types (masses and calcifications) were selected for each of the MUSICA and PLAHE algorithms and applied to all cases in the reader study. There have been changes in the various digital mammography systems tested here since this study was completed including: detector changes, hardware changes, and software changes that have led to more consistent image quality over earlier system versions. The Trex system tested here was Food and Drug Administration–approved in 2001 under the name Lorad by Hologic, Inc. (Bedford, MA), but was never marketed. The Fischer SenoScan tested here has undergone a number of software changes throughout the years. In November 2003, Fischer introduced a new image-processing algorithm that was designed not only to improve the overall image presentation quality compared with older Fischer default algorithms, but also to improve the speed of loading images into the review station for review. The GE Senographe 2000D has also undergone changes in display software and now offers a second image-processing algorithm, Premium View, in addition to the default algorithm at the review workstation. Whether any of these changes results in earlier detection of breast cancer by radiologists over previous versions of a given manufacturer’s digital mammography system or over screen film mammography would require additional reader studies. Although image processing alone was not a main effect of digital mammography versus screen film mammography in our study, it does seem to have an effect when combined with lesion and machine type. Training in the visual presentation of lesion characteristics when a given image processing algorithm is applied to a particular digital mammography system should be provided during applications training for digital mammography review workstations. Based on our results, multiple image-processing algorithms may need to be incorporated into future soft copy workstations to allow for optimal viewing of specific lesion types for each of the digital mammography systems. ACKNOWLEDGMENTS

The authors would like to acknowledge the contributions of the International Digital Mammography Development Group, especially the principal investigators of the project (3) that led to the creation of the image archive from which the raw images used in this study were obtained: Emily Conant, MD, University of Pennsylvania Medical Center, Philadelphia, PA; Laurie Fajardo, MD, University of Iowa, Iowa

IMAGE PROCESSING AND DIGITAL MAMMOGRAPHY

City, Iowa; Stephen Feig, MD, Mt. Sinai School of Medicine, New York, NY; Brad Hemminger, PhD, University of North Carolina, Chapel Hill, NC; Roberta Jong, MD, FRCPC-Sunnybrook & Women’s Health Science Center, Toronto, Ontario, Canada; Daniel Kopans, MD, Massachusetts General Hospital, Boston, MA; Andrew Maidment, PhD, University of Pennsylvania Medical Center, Philadelphia, PA; Bahjat Qaqish, PhD, University of North Carolina, Chapel Hill, NC; Rene Shumak, MD, Ontario Breast Screening Program, Toronto, Ontario, Canada; Melinda Staiger, MD, Monmouth Medical Center, Long Branch, NJ; Mark Williams, PhD, University of Virginia, Charlottesville, VA; and Martin Yaffe, PhD, Sunnybrook & Women’s Health Science Center, Toronto, Ontario, Canada. REFERENCES 1. Shah AJ, Wang J, Yamada T, et al. Digital mammography: a review of technical development and clinical applications. Clin Breast Cancer 2003; 4:63–70. 2. Pisano ED, Cole EB, Hemminger BM, et al. Image processing algorithms for digital mammography: a pictorial essay. Radiographics 2000; 20:1479 –1491. 3. Pisano ED, Cole EB, Major S, et al. Radiologists’ preferences for digital mammographic display. The International Digital Mammography Development Group. Radiology 2000; 216:820 – 830. 4. Baydush AH, Floyd CE, Jr. Improved image quality in digital mammography with image processing. Med Phys 2000; 27:1503–1508. 5. Chakrabarti K, Thomas JA, Kaczmarek RV, et al. Optimization of viewing conditions and phantom image quality evaluations on GE DMR and full-field digital mammography system. J Digit Imaging 2000; 13(Suppl 1):226 –227. 6. Qian W, Li L, Clarke L, et al. Digital mammography: comparison of adaptive and nonadaptive CAD methods for mass detection. Acad Radiol 1999; 6:471– 480. 7. Nawano S, Murakami K, Moriyama N, et al. Computer-aided diagnosis in full digital mammography. Invest Radiol 1999; 34:310 –316. 8. Cole EB, Pisano ED, Kistner EO, et al. Diagnostic accuracy of digital mammography in patients with dense breasts who underwent problemsolving mammography: effects of image processing and lesion type. Radiology 2003; 226:153–160. 9. Hemminger BM. Soft copy display requirements for digital mammography. J Digit Imaging 2003; 16:292–305. 10. Chan HP, Vyborny CJ, MacMahon H, et al. Digital mammography. ROC studies of the effects of pixel size and unsharp-mask filtering on the detection of subtle microcalcifications. Invest Radiol 1987; 22:581– 589. 11. De Maeseneer M, Beeckman P, Osteaux M, et al. Detecting clustered microcalcifications in the female breast: secondary digitized images versus mammograms. J Belge Radiol 1992; 75:173–178. 12. Karssemeijer N, Frieling JT, Hendriks JH. Spatial resolution in digital mammography. Invest Radiol 1993; 28:413– 419. 13. Feig SA, Yaffe MJ. Digital mammography. Radiographics 1998; 18:893–901. 14. Barlow WE, Lehman CD, Zheng Y, et. al. Performance of diagnostic mammography for women with signs or symptoms of breast cancer. J Natl Cancer Inst 2002; 94:1151–1159. 15. Kuzmiak CM, Millnamow GA, Qaqish B, et al. Comparison of full-field digital mammography to screen-film mammography with respect to diagnostic accuracy of lesion characterization in breast tissue biopsy specimens. Acad Radiol 2002; 9:1378 –1382. 16. Lewin JM, Hendrick RE, D’Orsi CJ, et al. Comparison of full-field digital mammography with screen-film mammography for cancer detection: results of 4,945 paired examinations. Radiology 2001; 218:873– 880.

595