Fig. 6
From: Explainable convolutional neural networks for assessing head and neck cancer histopathology
Exemplary test tiles with a negative correlation between Grad-CAMs generated for the classification network and segmentation network. For each pair of Grad-CAMs, the Pearson correlation coefficient (Corr) was computed. For classification, the predicted tumor probability (p) and for segmentation, the fraction of predicted tumor pixels (p) is given. On the top, the binary ground truth labels, i.e. the predominant class in the tile based on the maximum pixel count, are shown