Perceived Sufficiency of Full-Field Digital Mammograms With and Without Irreversible Image Data Compression for Comparison with Next-Year Mammograms

This study was conducted at a dedicated breast screening center at which 160 to 180 FFDM examinations are performed per day and more than 80,000 screening mammographies take place per year. To participate in the study, radiologists were required to have read a minimum of 2,000 mammograms per year for each of the past 3 years or to have recently completed a 1-year fellowship in which they read at least 10,000 mammograms. Because 12 radiologists were required for the study, 16 radiologists associated were recruited, and they signed a reader-informed consent agreement. The study protocol was approved by the Western Institutional Review Board in Olympia, Washington, and the study was conducted in accordance with the ethical principles of the Declaration of Helsinki.

For a brief orientation preceding the study session, eight sample cases (comprised of cancer and noncancer patients) were chosen from images not included in this investigation. Orientation included a review of the study case report forms (CRFs) and a demonstration of the PACS workstation (Synapse, FUJIFILM Medical Systems USA, Inc) on which the study was conducted.

In the study, the reading of digital mammograms took place in a totally darkened room dedicated to reading mammograms. To maintain confidentiality, patient identifiers on each digital image were limited to a unique study number and the subject’s initials. For images containing lesions, radiologists were not told either the type of lesion or the relative position. The order in which each reader evaluated the mammography images was staggered.

Image pairs were displayed on two high-resolution (5-megapixel) grayscale portrait monitors (Coronis 5MP Mammo, BarcoView LLC, Duluth, Georgia) calibrated to the Digital Imaging and Communications in Medicine (DICOM) Part 14 Display Standard. Luminance was set at 600 cd/m². Adjustments of the width and level of the image display window were not permitted. The display window width and center were set at 1,024/511, respectively. On-screen magnification up to one time was permitted (where 1 pixel displayed was equal to 1 pixel acquired). To ensure consistency and accuracy, this level of magnification was enabled by means of a hotkey on the PACS workstation.

The AON compression algorithm consists of a patented set of software tools for image compression, storage, and delivery. After a digital image is acquired, the AON compression engine, residing on the PACS server, will compress the image at multiple ratios depending on intended use and network capabilities, creating up to three versions of the acquired image. The compression ratios are configurable by image type, modality manufacturer, body part, and facility preferences. Images can be lossless-compressed or irreversibly (lossy) compressed at ratios up to 100:1. The AON compression engine writes a lossless-compressed DICOM Part 10 image version as the “original” image in order to meet the guidelines of the US Mammography Quality Standards Act (MQSA). The PACS user can then select which version of the image is to be delivered to the user workstation, the choice of which version depending on clinical requirements, available bandwidth, and other factors. When a compressed image is requested by a PACS user, the file is quickly delivered, and display software on the PACS workstation decompresses the image for viewing.

For the current study, two irreversible compression values were chosen after laboratory testing was conducted using the American College of Radiography (ACR) mammography accreditation phantom (MAP) to identify maximum data compression ratios meeting the minimum scoring criteria for phantoms as defined by the MQSA. Four different irreversible compression ratios (10:1, 20:1, 40:1, and 50:1) were tested in order to select a compression ratio at ≤20:1 and another compression ratio at >20:1 for the study. A certified radiographic technologist made two exposures of the MAP, one with automatic exposure control technique and a second with the mAs readout reduced to one-third. The images were sent to a PACS workstation as lossless-compressed images and as irreversibly compressed images in the four different compression ratios. Pairs of the phantom images were then analyzed by the investigator on two 5-megapixel DOME monitors by the standard procedure of counting the number of visible objects (e.g., fibers, masses, and microcalcifications). The irreversible compression ratios of 10:1, 20:1, and 40:1 met MQSA minimum scoring criteria as judged by the investigator; the irreversible compression ratio of 50:1 did not. For this study, the irreversible compression ratios of 20:1 and 40:1 were selected as representative of the PACS storage strategy for which the AON compression algorithm was created.

A total of 100 breast cancer screening or diagnostic cases were selected from an existing library of digital mammograms acquired by means of a Fujifilm Digital Mammography System (FUJIFILM Medical Systems USA) either at the University of California at Los Angeles, Los Angeles, CA, or at the Mayo Clinic, Rochester, MN. The cases were acquired at these facilities under patient-informed consent, and they were used previously by FUJIFILM Medical Systems USA, Inc, for submission of a premarketing approval application. The cases were drawn from patients with known true clinical status and with complete screen-film and digital mammography examinations (four standard views), in which there was sufficient anatomical coverage, sufficient contrast, no significant motion or other artifacts, no over- or underexposure of film, limited noise, and clinically insignificant difference in subject positioning between screen-film and digital mammography as determined by the image acquisition sites and the adjudicator. The 100 cases were chosen to include 50 tissue-proven cancers and 50 noncancers, with the diagnoses of noncancers determined by tissue sampling, mammography special views, ultrasound, or 1-year follow-up. Of the 100 cases, 17 were negative with 1-year follow-up. The 100 cases were also chosen to include 50 mediolateral oblique views and 50 craniocaudal views (Table 1). A further selection criterion was that the images represent a distribution of cancers and noncancers from (1) both screening and diagnostic populations, (2) all ACR Breast Imaging Reporting and Database System (BI-RADS) categories, (3) various types of main findings, and (4) all breast tissue composition categories, including cases with heterogeneously dense or extremely dense breast tissue (Table 1).

Study images were shown in pairs, with one image displayed per monitor, in three comparison modes: lossless-compressed images vs 20:1 irreversibly compressed images (mode 1), lossless-compressed images vs 40:1 irreversibly compressed images (mode 2), and 20:1 irreversibly compressed images vs 40:1 irreversibly compressed images (mode 3). Each reader rated all images in all three modes independently of the other readers. To ensure intrareader reproducibility, a subset of 20 cases were read twice in each of the three modes. Overall, then, each reader performed 120 reviews per comparison mode (Table 2). Compression levels were randomly assigned between the two monitors, and the order of presentation of image pairs was randomly assigned to counterbalance the order of presentation across modes (Table 3).

Readers were asked to record on the CRFs whether the quality of each right-side monitor image and left-side monitor image was sufficient for comparison with next-year mammograms and which image (right or left) had the lower level of data compression. In addition, if both images were perceived as sufficient for comparison with next-year mammograms, the readers were asked about their preference between the images on a 7-point Likert scale ranging from +3 (image on the right is much better) to 0 (similar) to –3 (image on the left is much better). Secondary end points similarly evaluated on a 7-point Likert scale were which image (right or left) better displayed an object if there was an area of interest on the image, which image (right or left) better displayed the area at or near the skin line, and which image (right or left) better displayed the area at or near the chest wall.

A total of 12 radiologists was considered adequate for statistical analysis. Because of potential for scheduling conflicts with the time of the study, 16 radiologists were recruited to participate, and all were able to do so. For one of the 16 radiologists, however, study image pairs were accidentally flipped on the monitors in mode 1 after software at the radiologist’s workstation was restarted when the initial sequences of image pairs did not load properly. After the software was restarted, images appeared on the two monitors in the software’s default setting, not the programmed study setting. The restarting of the workstation’s software was noted by the study monitors, and the setup anomaly was afterward verified. Because of the accidental software error, data from this radiologist were excluded from the study. Data from the other 15 radiologists were used for analysis.

The primary objective of the study was to demonstrate the noninferiority of irreversibly compressed images vs lossless-compressed images in terms of perceived sufficiency of image quality for comparison with next-year mammograms. A random-effect analysis was applied to ascertain the difference between compression level pairs in each mode in terms of the per-reader average proportion of images perceived to be of sufficient quality for comparison with next-year mammograms. The Obuchowski model was used to calculate sample sizes and to compute the standard error of the per-reader average for each compression level in each mode.8,9 The Student’s t distribution was used to obtain confidence intervals (CIs). Although the Obuchowski model is most commonly applied to the area under the receiver operating characteristic (ROC) curve, it can by extension be applied to proportions of successes. The end point of perceived sufficiency of image quality for comparison with next-year mammograms, it should be noted, is not the same as an end point of diagnostic accuracy, for which an ROC study would have been appropriate. The irreversibly compressed images in this study are not intended for detecting or diagnosing malignancies but for making decisions about next-year mammograms.

The random-effect analysis permitted generalization of perceived sufficiency beyond the number of cases and readers in this study. Within each mode, noninferiority was considered demonstrated if the lower limit of the 98.33% CI for the average difference between levels of compression in terms of the per-reader average proportion of images perceived to be of sufficient quality exceeded the value of −0.05. The noninferiority delta was set at 5%. At an alpha level of 0.05/3, a target sample size of 100 comparison cases with 12 readers was calculated.

For overall preferences between the paired images and for preferences in terms of object display and display of areas at or near the skin line and at or near the chest wall, the alpha level was set at 0.5/3 to protect against multiple comparisons. For each comparison mode, the per-reader average proportion of scores that were at least 0 (indicating that the more compressed image has image quality similar to or better than the less compressed image) was calculated, along with corresponding 98.33% CIs.

For reader ranking of image compression, the alpha level was set at 0.05/3 to protect against multiple comparisons. For each mode, the average proportion of correct responses was calculated along with corresponding 98.33% CIs. A 98.33% CI was computed using a normal approximation and a standard error adjusted for correlations among the proportions. If the CI was <0.50, it was concluded that the images were comparable, since readers correctly identified more compressed images less often than would be expected by chance.

Intrareader reproducibility was also examined for all end points using data from the 20 cases in the subset. For each reader, in each mode, proportions of observed agreement and of expected agreement were computed. Kappa statistics were computed when the proportion of expected agreement was less than perfect.

Readers dictated responses to transcribers, who filled in supplied CRFs by pen. The CRFs were managed by PharmaPro Corporation, Inc. (Cambridge, MA). Data analysis was performed by Biostatistics Consulting, LLC (Toronto, Ontario, Canada).