Handling multiple classes in SCGM Spinal Cord Challenge Dataset

[zahilshanis]

Hi @perone , Thank you very much for sharing the code of your work.

I am working on a domain adaptation research project and am trying to reproduce the results of your work for comparison. I see that SCGM Spinal Cord Challenge Dataset has 3 labels ( background, grey matter and white matter) and the challenge is to segment grey matter region from other regions. Your published code (main.py) is not explicitly transforming ground truth labels from 3 classes to binary class and uses medicaltorch's SCGMChallenge2DTrain class for slicing and data loading, which assumes that the input is of binary class.

I converted ground truth label 2 (white matter) to 0 and trained it as a binary class problem for grey matter segmentation using your code. Unfortunately, I am getting totally different results for the experiments listed in your paper. For the supervised training using centers 1 and 2 with no domain adaptation, I'm getting 87.4, 89.27, 5.39 and 39.02 as the dice coefficients for centers 1, 2, 3, and 4 respectively compared to 47,25, 50.69, 82.81 and 69.41 you recorded for the same. It looks like I am missing something in the data processing pipeline before training the network. If you are doing any additional data preprocessing or transformation before training, could you please share the details of the same?

Thanks