Guiding Ultrasound Breast Tumor Classification with Human-Specified Regions of Interest: A Differentiable Class Activation Map Approach

Ultrasound is a useful tool for evaluating the malignancy of breast tumors, but the interpretability of deep learning models used for benign-versus-malignant classification remains a challenge. Class Activation Maps (CAMs) offer a method to visualize and understand the neural network (NN) focus, but there lacks an established method to correct erroneous foci by the NN. To address this gap, we developed a deep learning framework that guides the NN's focus using human expert-specified Regions of Interest (ROIs). By formulating focus correction as a distance minimization task and introducing a differentiable peaked CAM (PCAM), our method ensures the NN's focus aligns with the human-specified ROI. By minimizing the difference between the PCAM and the binary ROI mask, the focus of NN is effectively driven into the ROI region. Evaluations on multi-center public datasets demonstrate the effectiveness of incorporating human-specified ROIs in improving both tumor classification accuracy and focus localization accuracy.