XWH - 05 - 1 - 0292 TITLE : A Computer - Aided Diagnosis System for Breast Cancer Combining Digital

Purpose: To develop computer-aided diagnosis (CADx) models using both mammographic andsonographic descriptors and to estimate the generalization performance of these models onfuture cases.Materials and Methods: Institutional Review Board approval was obtained for this HIPPA-compliant study. Mammographic and sonographic exams were performed on 737 patients,yielding 803 breast mass lesions (296 malignant, 507 benign). Radiologist-interpreted featuresfrom the mammograms and sonograms were used as input features by a linear discriminantanalysis (LDA) and an artificial neural network (ANN) to differentiate benign from malignant lesions. An LDA using all the features was compared to an LDA using only stepwise-selectedfeatures. Classification performances were quantified using receiver operating characteristic(ROC) analysis and were evaluated in a train, validate, and retest scheme. On the retest set, bothLDAs were compared to the radiologists overall assessment score of malignancy.Results: Both the LDA and ANN achieved high classification performance with cross-validation(AUC = 0.92 ± 0.01 and 0.90AUC = 0.54 ± 0.08 for the LDA, AUC = 0.92 ± 0.01 and 0.90AUC = 0.55± 0.08 for the ANN). Both models also generalized very well to the re-test set, with no statistically significant performance differences between the validate and retest sets (p > 0.1). On the retestset, there were also no statistically significant performance differences between the LDA using allfeatures and using only the stepwise selected features (p > 0.3) and between either LDA and theradiologists assessment score (p > 0.2).Conclusion: The results showed that combining mammographic and sonographic descriptors ina CADx model can result in high classification and generalization performance. On the retest set,the LDA matched the radiologists classification performance.