Using this for creating backpropagation saliency maps?

[greglahmer]

If we are trying to establish, just the pixels instead of image regions with the highest gradients attached to them, I have heard saliency maps using backpropagation is the best way? Is it possible to do this with medcam as well?

[Karol-G]

You probably mean Grad-CAM? This is the default setting of medcam.

[stevenagl12]

I think they may be referring to saliency maps that calculate the weight of the pixels or voxels in an image, not the gradients. Maybe something like: https://towardsdatascience.com/saliency-map-using-pytorch-68270fe45e80.

[greglahmer]

Yes, that is what I meant. The saliency maps for calculating the weight. I am not sure how to do that for segmentations in 3D, so I thought the package might be able to do it...