Keynote Lecture — Breast Cancer
Monday 9 October 2000, 17.00–17.30
Cancer Imaging volume 1, pages 25–27 (2000)
Computer-aided detection (CAD) in screening mammography
Interpretation of radiological (and other complex medical) images is challenging due to an often overwhelming amount of visual data, much of which is of no consequence and which often serves to confound the observer. In addition, there are well-recognized physiologic obstacles to visual search and recognition, which contribute to false-negative radiological interpretations. The interpretation of screening mammograms is particularly challenging, because a large number of cases are viewed to detect a small number of cancers (3–10 cancers per 1000 women screened), which are often manifest by subtle alterations superimposed upon the complex radiographic structure of the breast.
To overcome known limitations of human observers, second (or double) readings of screening mammograms have been implemented in some practices, which yields a 4–15% increase in cancers detected. For two decades, researchers have investigated the use of computers to analyse mammograms to provide prompts to the radiologist to minimize oversight errors (observational lapses).
In 1998, the FDA approved the use of a computer aided detection (CAD) system for screening mammography. This system is specifically designed to minimize the possibility of oversights leading to false-negative interpretations by pointing out areas on the mammograms that have features suggestive of cancer. The system digitizes the mammogram at 50-micron spatial resolution and 12 bits of gray scale, with excellent linearity and low noise over the optical density range of interest. The digitized image is analysed by proprietary signal processing software, which highlight ROIs (regions of interest) with the following characteristics:
clusters of bright spots (i.e. regions suggestive of microcalcification clusters), marked by triangles; and, dense regions and dense regions with radiating lines (i.e. regions suggestive of masses or architectural distortions) marked by asterisks.
The end result of the analysis is a low-resolution image of the original mammogram, with the detected ROIs marked by triangles and asterisks which are then superimposed on that image to allow the radiologist to locate the areas on the original films. The system marks particular ROIs because of their physical characteristics. That is, the marked ROIs are associated with a visually perceptible structure that has some of the generally accepted geometric characteristics of microcalcifications or masses. The marked areas may be something other than an actual abnormality, but these are readily recognized as such by the radiologist upon review of the original mammogram.
Three comprehensive clinical studies were done for the FDA approval process, the results of which were recently published.
Study 1: How many cancers are missed (false negative readings) by the current screening mammography process; and, can a CAD device detect and mark these cancers?
All breast cancer cases diagnosed at 13 clinical sites by screening mammography (n = 1083, Current Mammograms) over a 2-year period, and the most recent available prior screening exam (n = 427, Prior Mammograms) acquired in the previous 9–24 months, were obtained.
In a ‘side-by-side’ comparison of the Current with the Prior Mammogram, and with knowledge of the cancer’s eventual location on the Current Mammogram, 67% (286/427) of breast cancers were visible in retrospect on the Prior mammogram. However, when a panel of five radiologists performed a blinded review of the Prior Mammograms to determine those that were ‘actionable’, the false-negative rate was 21%.
The performance of the CAD system was evaluated on each of the Prior Mammograms that were identified as actionable by the majority of the Panel Radiologists. The results of the blinded review of the Panel Radiologists and CAD system performance are below:
The original radiologists’ sensitivity in detecting asymptomatic breast cancer was calculated as 79%. The capability of the CAD system to mark abnormalities correctly in the Prior Mammograms considered actionable by the Panel Radiologists demonstrates the effectiveness of the device in minimizing the possibility of false-negative readings due to observational lapses during mammographic screening.
Study 2: Does the use of CAD increase the radiologist’s diagnostic workup rate?
Data on workup rates from 14 participating radiologists at five clinical sites before (23 682 cases) and after (14 817 cases) the installation of a CAD system showed no increase in the work-up rate for any individual radiologist or for the group overall. The overall preinstallation work-up rate was 8.3% compared with the post-installation work-up rate of 7.6%.
Study 3: Can a CAD device mark regions of interest associated with biopsy proven cancer; and, what is the intrasystem and intersystem precision of marking these regions?
All available consecutive screening mammograms on which the cancer was diagnosed (n = 1083, Current Mammograms) were obtained from 13 clinical sites. CAD analysis of these cases was performed. The CAD system was scored to have marked the case correctly if it identified the correct location and classification of the biopsy-proven lesion.
Twenty-five Current Mammograms, on which the cancer had strong features and was seen on both views, were processed by the CAD system to evaluate the intrasystem and intersystem reliability in marking characteristics associated with cancer. The two views were run 10 times on three different CAD systems. The CAD system was scored to have marked the case correctly if it identified the correct location and classification of the biopsy-proven lesion.
Improvements in algorithms
Continued refinements of algorithm design are being made to provide increased sensitivity for cancer detection while diminishing the false marker rate. The above data were generated with algorithms (version 1.2) used in the clinical trials for the PMA submitted for FDA approval over 2 years ago. Data on the same case sets (Current and Prior Mammograms) for a new, recently FDA approved (April 2000) algorithm (version 2.2) will be presented, which demonstrate improvements in the mass code and a further 20% decrease in the false marker rate.
Temporal and right/left comparison
Current algorithms analyse breast tissue for evidence of features associated with cancer, much as the radiologist does. However, radiologists in addition compare the current exam to prior studies for evidence of interval change (emerging densities) and compare the right and left breasts for evidence of asymmetries (asymmetrical densities). Incorporation of similar analytic strategies would particularly enhance detection of those cancers with mass features.
Associating the ‘likelihood’ of cancer with differentiating markers
Efforts are being made to emphasize those markers that point out features with a ‘higher probability’ of representing breast cancers. This assignment of ‘probability’ by specific markers should prompt the reader to pay particular attention to these flagged feature sets.
Improvements in the display system
The current CRT display marks the centroid of areas associated with features suggestive of microcalcifications and masses (densities with or without spiculated features) on each view of the breast. Enhanced display capability will allow precise mapping of these features (highlighting each suspect microcalcification and outlining the suspect mass) as well as defining the same feature (or area) on the orthogonal view. This should facilitate radiologist viewing and interpretation time.
Integration of CAD with full field digital mammography (FFDM)
Incorporation of CAD with FFDM will simplify the current process, by eliminating the need to first digitize the radiographs and then to devise an independent display medium for the CAD results. Thus, the CAD algorithms will directly interact with the digitally obtained image and the display of the CAD results can be incorporated into the viewing CRT (with soft copy interpretation). As, or perhaps more importantly, CAD can greatly facilitate the viewing of the digitally derived images by pre-selecting the appropriate ‘window and level’ settings to present the salient information to the radiologist.
CAD technology is a highly robust system with excellent sensitivity for the identification of microcalcification and, to a lesser extent, features associated with cancer on screening mammograms. Use of the CAD system did not result in any statistically significant increase in patient work-up rates. Among cases defined as actionable by 4/5 or 5/5 Panel Radiologists that were acquired 9–24 months before the cancer was actually detected, the CAD system correctly identified features suggestive of cancer in 95% of cases.
To date, over 100 CAD units have been installed at screening mammography sites in the USA and over 700 000 women have had their screening mammogram interpreted with CAD assistance.
With the newest algorithms (version 2.2), the CAD system demonstrates:
99% detection rate for cancers displaying microcalcifications
86% detection rate for cancers displaying mass features
90% detection rate for all breast cancers
Importantly, for every 100 000 breast cancers currently detected by screening mammography, use of CAD could result in the detection of an additional 20 500 cancers.
CAD can play an important role in assisting the radiologist to minimize observational lapses by suggesting areas on the original mammogram that may warrant a second review. The contributions of CAD, which is in its infancy, to complex image interpretation in mammography and in all fields of medicine is likely to be significant.
The CAD system used in these studies is a product of R2 Technology, Los Altos, CA, of which Dr Castellino is the Chief Medical Officer.
Burhenne LJ, Wood SA, D’Orsi CJ, Feig SA, Kopans DB, O’Shaunessy KF, Sickles EA, Tabar I, Vyborny CJ, Castellino RA. The potential contribution of computer-aided detection to the sensitivity of screening mammography. Radiology 2000; 215: 554–62.
About this article
Cite this article
Keynote Lecture — Breast Cancer. cancer imaging 1, 25–27 (2000). https://doi.org/10.1102/1470-7330/00/010025+03