Contributor: Junbo Shen
Contribution Type: Interpretability method, tests, example

Description

Add GradCAM to pyhealth.interpret.methods for CNN-based medical
image classification.

Main behavior:

• supports target_layer as either an nn.Module or a dotted path such as
  model.layer4.1.conv2
• returns image heatmaps as {input_key: cam} with shape [B, H, W]
• uses the predicted class by default when class_index is not provided
• uses logit when available and falls back to y_prob otherwise

And add:

• chest X-ray tutorial script using Grad-CAM; API docs for Grad-CAM; unit tests for Grad-CAM

Files added/changed

1. pyhealth/interpret/methods/gradcam.py
2. tests/core/test_gradcam.py
3. examples/cxr/gradcam_cxr_tutorial.py
4. docs/api/interpret.rst
5. docs/api/interpret/pyhealth.interpret.methods.gradcam.rst
6. docs/tutorials.rst

Testing

Validated locally with:

python -m py_compile \
  pyhealth/interpret/methods/gradcam.py \
  tests/core/test_gradcam.py \
  examples/cxr/gradcam_cxr_tutorial.py

python -m unittest tests.core.test_gradcam -v

python examples/cxr/gradcam_cxr_tutorial.py \
  --root /Users/junboshen/Documents/COVID-19_Radiography_Dataset \
  --weights none \
  --output /tmp/gradcam_cxr_overlay.png

The Grad-CAM implementation was also checked against the original
jacobgil/pytorch-grad-cam implementation by comparing the output.