Table 1 Names and output sizes of layers, from which feature maps were obtained. HR-CAM uses feature maps from several layers, whereas Grad-CAM only requires the last convolutional layer’s output. The layer names correspond to the original layer names of EfficientNet-B0 [20]
From: Explainable convolutional neural networks for assessing head and neck cancer histopathology
Method | Classification CNN layers | Segmentation CNN layers | Output size [px] |
|---|---|---|---|
HR-CAM | block_3a_expand_activation | block_3a_expand_activation | 56\(\times\)56 |
block_4a_expand_activation | block_4a_expand_activation | 28\(\times\)28 | |
block_6a_expand_activation | block_6a_expand_activation | 14\(\times\)14 | |
block_7a_expand_activation | block_6d_expand_activation | 7\(\times\)7 | |
Grad-CAM | top_activation | block_7a_expand_activation | 7\(\times\)7 |