Classifying presence or absence of calcifications on mammography using generative contribution mapping

The purpose of this study was to verify the efficacy of generative contribution mapping (GCM), an explainable deep learning model for images, in classifying the presence or absence of calcifications on mammography. The learning dataset consisted of 303 full-field digital mammography (FFDM) images labeled with microcalcifications obtained from the public INbreast database without extremely dense images. FFDM images were divided into calcification and non-calcification patch images using a sliding window method with 25% overlap. The patch images of the mediolateral oblique (MLO) and craniocaudal (CC) views were divided into a training set of 70%, a validation set of 10%, and a testing set of 20%. The classification performance of GCM classifiers was evaluated and compared with that of EfficientNet classifiers. Visualization maps of GCM highlighted regions of interest more clearly than EfficientNet's gradient-weighted class activation maps. The results showed that GCM classifiers yielded an accuracy of 0.92 (CC), 0.91 (MLO), and an area under the receiver operating characteristic curve of 0.92 (CC), 0.94 (MLO). In conclusion, GCM could accurately classify the presence or absence of calcifications on mammograms and explain intuitively reasonable grounds for their classification with visualization maps highlighting regions of interest.