Abstract
Background/Objectives: Breast cancer is one of the most common cancer types in women, and reading a breast ultrasound is still very much dependent on the individual radiologist, creating inter-observer variability. Deep-learning classifiers can achieve accuracy comparable to that of experts, but there are two challenges for their use in routine practice: they rarely explain their predictions, and a model can reach the right label for the wrong reason, by attending to tissue unrelated to the lesion. This study addresses both of these. Methods: On the BUS-BRA dataset, we used the lesion segmentation mask as an additional input channel to the network, along with the RGB image, so two backbones (EfficientNet-B3 and EfficientNet-B4) were directed towards the lesion from the first layer, while only the initial convolution layer was modified, keeping the ImageNet-pretrained weights unchanged. To ensure a fair evaluation, the data were split at the patient level rather than the image level, preventing leakage. Training was performed with focal loss, heavy augmentation and test-time augmentation, with the decision threshold set on the validation set. Grad-CAM, SHAP, and Occlusion Sensitivity explained each prediction, and were scored for fidelity, stability, and overlap with the true lesion. Results: During the controlled ablation, the mask channel improved accuracy by more than eight points over an region-of-interest (ROI)-only baseline. The average of the two backbone models yielded 86.7% accuracy and 0.911 AUC on the held-out test set, and when applied to the independent BUSI dataset, without any modification, with the threshold frozen, it achieved 90.4% accuracy and 0.972 AUC, with Grad-CAM maps landing on the annotated lesions. Conclusions: The accuracy of mask guidance using an imbalance-aware design is competitive, and the accuracy can be transferred to an external dataset, providing a credible basis for decision support that can be verified by the clinician with the image.

This publication has 49 references indexed in Scilit: