Figure 3.
Post-hoc analysis using the trained SG and PANDA models
(A) Attribution maps obtained using integrated gradients32 with blurred images as baselines are presented for a malignant sample in the test set of SG gland classification dataset. This sample is predicted correctly by the SG and PANDA models. See also Figures S3 and S4.
(B) One malignant patch and one benign patch with the highest correct class probability scores (by the SG model) from nine different slides in the test set of the SG gland classification dataset. Predicted malignancy scores by the trained SG and PANDA models are presented under each image. See also Figure S5. Scale bars are shown in black.