Figure 5
Representative, correctly-classified cancer image-patches from each dataset, visualized with cancer and normal components using the grad-CAM technique41. The gradients were traced from the last convolutional layer before the inception modules to the cancer and normal logits separately to visualize contributions from cancer and normal features leading to a correct cancer prediction. (a) Breast cancer SLN metastasis image patch correctly classified as cancer with 58% probability. (b) Papillary thyroid carcinoma image patch correctly classified as cancer with 89% probability. (c) SCC image patch correctly classified as cancer with 99% probability.
