Application of a computer-aided detection (CAD) system to digitalized mammograms for identifying microcalcifications

PURPOSE: It is estimated that during mammographic screening programs radiologists fail to detect approximately 25% of breast cancers visible on retrospective review; this percentage rises to 50% if minimal signs are considered. Independent double reading is now strongly recommended as it allows to reduce the rate of false negative examinations by 5-15%. Recent technological progress has allowed to develop a number of computer-aided detection (CAD) systems. The aim of CAD is to help radiologists interpret lesions by serving as a second reader. In this study the authors developed and applied a CAD system to measure its ability to microcalcifications detect and compare its performance with that of a human observer. MATERIAL AND METHODS: The study was performed as part of the CALMA (computer-aided library for mammography) project of the Pisa section of the National Institute of Nuclear Physics. The aim of this project is to set up a large database of digital mammograms and to develop a CAD system. Our study series consisted of 802 mammograms - corresponding to 213 patients - digitalized between March and June 2000. We performed traditional mammography and then digitalized the mammograms using a CCD linear scanner (pixel size of 85 x 85 microm2, 12 bits). The images were evaluated by two radiologists with similar experience and then by the CAD system. This CAD system searches for microcalcifications by using ad hoc algorithms and an artificial neural network (Sanger type). RESULTS: The number of clusters in our database was 141 corresponding to 140 images; 692 images were non pathological. The CAD system identified a variable number of clusters depending on the threshold values. The threshold value is a number over which the probability of finding a lesion is highest. With thresholds of 0.13 and 0.14 the CAD system identified 140/141 clusters (99.3%); with a threshold of 0.15 it identified 139/141 clusters (98.6%); with a threshold of 0.16, 137/141 (97.2%); with a threshold of 0.18, 133/141 (94.3%); with thresholds of 0.18 and 0.20, 130/141 (92.2%). With threshold values of 0.13, 0.14, 0.15, 0.16 and 0.17 the system's sensitivity was greater than 82%, whereas with values of 0.18 and 0.20 it was greater than 80%. The number of false positive region of interest (ROI) / image was greater with low threshold values: in particular, thresholds of 0.13 and 0.14 yielded 16 false positives /image, thresholds of 0.15 and 0.16 yielded 9 and 7 false positives/image, and both 0.18 and 0.20 only 5/image. DISCUSSION: ROC curve shows how the use of high threshold values determined a very high specificity despite very low sensitivity rates. Conversely, low threshold values allowed to have a high sensitivity and a very low specificity. The best performance of our CAD system was obtained with threshold values at 0.15 and 0.16. In fact these thresholds resulted in a high sensitivity (greater than 82%) with an acceptable number of false positives/image, 9 and 7/image, respectively. It is not yet known how radiologists can deal with large numbers of false positives in screening programmes but in our opinion the most important feature of a good CAD system is a high sensitivity. CONCLUSIONS: In the near future the use of CAD systems will be widespread and easier to apply to everyday practice above all in centers where digital mammography is performed. Mammograms could be directly shown to radiologists after the CAD system has selected the ROI and analysed the images. Thanks to its high sensitivity and despite its low specificity CAD represents a concrete aid for radiologists.