Detection of breast carcinomas potentially missed during screening by means of a standalone CAD system

PURPOSE: To develop and evaluate a standalone computer-aided detection (CAD) system for the identification of suspicious non-referred cases. These cases could be submitted to a third reader for reconsideration. METHOD AND MATERIALS: We investigated a CAD system for mass detection consisting of three stages: pre-processing, initial detection and a final classification. In the pre-processing stage, the input image is segmented and enhanced. Next, suspicious regions are detected by an ensemble of five neural networks provided with five features based on spiculation and gradient measures. In the final stage, these candidate regions are assigned a malignancy score based on a richer set of 73 features that, in addition, measure contrast, location, linear texture, etc. For this final stage three different classifiers were evaluated: linear discriminant analysis (LDA), gentleboost with regression stumps and support vector machines (SVMs). The classifier parameters were optimized at a very low recall rate (2%), in contrast to previously proposed CAD systems, because we are interested in such a high specificity in order to minimize the reviewing effort for a possible third reader. The data for this study consisted of a set of 28,811 scanned film mammograms collected from different sites in the Netherlands. Of these, 740 images correspond to 266 malignant cases with visible masses and architectural distortions that were not detected until the next screening round (two years later). The remaining images correspond to 9,127 normal cases with no sign of pathology. A five fold cross-validation scheme was used to evaluate the system. Performance was assessed by mean case sensitivity obtained at 2% recall rate. Comparisons were made using bootstrap analysis. RESULTS: At a 2% additional recall rate, mean sensitivities achieved by LDA, gentleboost and SVM were 4.0%, 9.8%, and 10.5%, respectively. SVM and gentleboost performed significantly better than LDA. Using the best classifier, the CAD system was able to detect 28 (10.5%) previously missed cases for reconsideration. CONCLUSION: Automatic selection of non-referred suspicious cases for reconsideration using a standalone CAD system is feasible at low additional recall rates. CLINICAL RELEVANCE/APPLICATION:Using a CAD system that autonomously detects potentially missed cancers could substantially improve screening efficacy with only a small additional reading effort.