Tutorial gradCAM 3D seems to have inverted heatmap

[CamilleDelgrange]

Bug - GradCAM:
https://github.com/Project-MONAI/tutorials/blob/main/modules/interpretability/class_lung_lesion.ipynb

Following this tutorial, and comparing with the original paper on GradCAM (Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization), I have the impression that there is an issue somewhere with the activation score or the heatmap colors because red should be related to high importance score and blue the reverse.

How can I fix that in my situation ?

[KumoLiu]

Hi @CamilleDelgrange, thanks for your interest here.
Yes, the default_normalizer used in GradCam will flip magnitudes.
https://github.com/Project-MONAI/MONAI/blob/dev/monai/visualize/class_activation_maps.py#L37

You can simply modify it by adding a customized normalizer.

def custom_normalizer(x):
    def _compute(data):
        scaler = ScaleIntensity(minv=0.0, maxv=1.0)
        return np.stack([scaler(i) for i in data], axis=0)
    if isinstance(x, torch.Tensor):
        return torch.as_tensor(_compute(x.detach().cpu().numpy()), device=x.device)
    return _compute(x)

monai.visualize.GradCAM(nn_module=model_3d, target_layers="class_layers.relu", postprocessing=custom_normalizer)

Hope it helps, thanks.