ConvNet Model Evaluation for Binary and Multilabel Breast Mammogram Mass Segmentation

Breast mass segmentation in mammograms is a vital step in breast cancer screening. Convolution neural networks (ConvNets) based models have been widely used to segment the breast mass from mammograms, which can be further classified into either benign or malignant. Recently, end-to-end frameworks have been proposed to perform breast mass segmentation and classification in one stage. In this study, we assess the performance of common binary and multi-label segmentation models designed for one-stage mass segmentation and classification tasks. It is important to note that previous investigations predominantly confined their training and testing to abnormal mammograms, those displaying mass abnormality. Furthermore, their evaluation metrics primarily revolved around the degree of spatial overlap between model predictions and ground truth, commonly quantified using the Dice score. However, in the context of a clinically feasible end-to-end framework, the model must be equipped to process both normal and abnormal mammograms as input, necessitating a more comprehensive evaluation strategy. Thus, In the presented work, we aimed to evaluate the binary and multilabel segmentation frameworks concerning end-to-end breast mass segmentation on a private dataset recently prepared by authors. For training, we employed U-Net variants, introducing the False Positive Rate (FPR) alongside the Dice score for performance evaluation. Furthermore, We used information regarding the class of masses (benign or malignant) and breast tissue density provided with each mammogram to analyze the segmentation model’s performance with respect to these parameters.