✨🔨 add distance transform

[dyollb]

What do these changes do?

• ✨this PR adds DistanceTransform transform to compute (boundary) distance. The boundary distance can be used to compute a boundary/distance loss.
• ✨ a BoundaryLoss was added, which uses pre-computed distances from the boundary of the ground truth masks to compute how dissimilar (far) the predicted masks are.
• 🔨 transforms were moved to a new submodule transforms

The boundary loss currently integrates (sum) over boundary deviation (XOR(A_gt, A_pred)), see b):

While the pre-computation is on the CPU (using scipy.ndimage.distance_transform_cdt), the loss is evaluated using simple operations using torch.Tensors.

Related issue/s

• partially implements Add boundary loss #58

How to test

The distance transform is tested in tests/transforms/test_distance.py.

The boundary loss should be tested

• on GPU
• and in use, i.e. run training

Checklist

• Non-breaking change (fix or new feature that would not break existing functionality).
• Unit tests for the changes and run locally with the command pytest tests
• Runs on minimum Python version

[codecov]

Codecov Report

Merging #57 (dc79f5f) into main (9e176d7) will increase coverage by 1.36%.
The diff coverage is 67.88%.

@@            Coverage Diff             @@
##             main      #57      +/-   ##
==========================================
+ Coverage   53.90%   55.27%   +1.36%     
==========================================
  Files          20       22       +2     
  Lines        1228     1328     +100     
  Branches      218      237      +19     
==========================================
+ Hits          662      734      +72     
- Misses        530      545      +15     
- Partials       36       49      +13     

Flag	Coverage Δ
unittests	55.27% <67.88%> (+1.36%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
src/segmantic/transforms/detect.py	74.21% <ø> (ø)
src/segmantic/transforms/ensemble.py	100.00% <ø> (ø)
src/segmantic/transforms/io.py	86.30% <ø> (ø)
src/segmantic/seg/monai_unet.py	20.77% <50.00%> (+0.12%)	⬆️
src/segmantic/seg/boundary_loss.py	52.08% <52.08%> (ø)
src/segmantic/transforms/distance.py	81.35% <81.35%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[Konohana0608]

Awesome. Thank you for this. Will try to use it as soon as the cross-validation with the new bone masks is done~~