neuropoly/medicaltorch

Potential merge issue with get_item method ?

Closed this issue · 1 comments

AnBucquet commented

Thanks to @andreanne-lemay, the dataset class MRI2DSegmentationDataset supports multi-channel input. However the __getitem__ method returns a single tensor (which is the concatenation of all the input tensors) in that case. I think that might be a bit too specific to our problem and might be a problem to merge our repo with the original MedicalTorch.

Would it be possible to return a list of tensors instead (even in the case of a singe channel) ? A stack transformation can then be applied to get the same output.

Additionally, the same philosophy might be applied to the GT and ROI as they are not necessarily identical regarding the modality (even if they should ...).

charleygros commented

The advantage that I see with your solution @AnBucquet is that it would allow to process multi-modal data either in a multi-channel way, either with CNNs in parallel (i.e. multi-path configurations: one branch processes T2w and another one, T2*w). Quite flexible!

  • 👍1