Ivadomed: Integration of medicaltorch in ivadomed

[andreanne-lemay]

Since the fork neuropoly/medicaltorch is starting to differ a lot from the original medicaltorch repository, the fork will be integrated to the ivado-medical-imaging repository. This will centralize the code and avoid to make changes to 2 different repositories.

Different steps for the medicaltorch integration:

• [x] Merge metrics (dice, recall, specificity, etc.)
• [x] Integrate filters
• [ ] Merge transforms. Currently medicaltorch transforms are parent to ivadomed transforms. Moving the transforms will remove this relation. Following class should be moved in ivadomed/transforms.py
  -UndoTransform
  -ToPIL
  -NormalizeInstance
  -NormalizeInstance3D
  -ToTensor
  -StackTensors
  -Crop2D
  -CenterCrop2D
  -ROICrop2D
  -RandomRotation
  -RandomRotation3D
  -RandomReverse3D
  -RandomAffine
  -ElasticTransform
  -Resample
• [ ] Move following classes in medicaltorch/datasets.py to ivadomed/loader.py
  -SegmentationPair2D
  -MRI2DSegmentationDataset
  -MRI3DSubVolumeSegmentationDataset
  -mt_collate (function)
  -SampleMetadata

The ivadomed/loader.py may become too dense. We might want to create a loader_utils.py.

[jcohenadad]

The ivadomed/loader.py may become too dense. We might want to create a loader_utils.py.

👍 anything that could reduce the code in loader (which is already long) is welcome

[charleygros]

I even wonder if we should create a folder loader with core.py, utils.py, adaptative.py etc.
Similarly a model folder with something like: core.py, unet_2d.py, unet_3d.py, hemis.py, film.py, utils.py. ... one day :)

[olix86]

I think we can close this, good job everyone :)