Hausdorff distance loss support #6993
Description
Hausdorff distance is widely used in evaluating medical image segmentation methods. Adding an objective/loss function directly to optimize this distance can be instrumental in optimizing this score [1].
An existing implementation is publicly available on GitHub by Patryk Rygiel, although it might take some engineering work to integrate within the MONAI framework.
I was able to train a 3D segmentation model successfully using MONAI's SwinUNETR with this implementation and I would like to contribute the code for this loss so that it's more widely available and easier to integrate with MONAI-based repositories.
References:
[1] Karimi, D., & Salcudean, S. E. (2019). Reducing the Hausdorff distance in medical image segmentation with convolutional neural networks. IEEE Transactions on medical imaging, 39(2), 499-513.