Improved Deep Learning Explanations for Prostate Lesion Classification through Grad-CAM and Saliency Map Fusion

Explainability techniques have been recently postulated as being effective for localizing medical regions of interest. However, there has been recent criticisms regarding the validity of these approaches in terms of robustness and accuracy. This study explores if combining two saliency methods, namely Saliency Maps and Grad-CAM, improves the robustness and accuracy of localization. The experiments are three-part. First, the accuracy of localization is measured and repeatability experiments are performed to determine if the results are repeatable across model architectures. Second, cascading randomization experiments are conducted to determine the robustness to changes in model parameters. Third, we measure the overlap between Saliency Maps and Grad-CAM and show there is a relationship between the overlap and accuracy of prediction.