Adding model to segment bones

[gattia]

Starting a pull request while working on this to get feedback. A few things are done differently than other models.

• Currently loading model + weights from .h5 file (vs. building model & loading weights)
  • This requires storing shape of original images used during training, reshaping inputs to that shape, segmenting, and then returning the shape. https://github.com/gattia/DOSMA/blob/adcebe3df0556eaddac95b3d7501d313f361a367/dosma/models/stanford_qdess_bone.py#L180-L215
  • Added option to "resample_images" which if True uses this strategy of using the original model (and resampling shape). If False, presumably we will re-build the original model and load the weights. This is essentially a placeholder for me to get code to re-generate the model(s) directly - I was worried at first I had something in the network that wasn't fully convolutional (Dense layer), but I think we are good. Right now this throws an error (will implement in future).
• Use argmax to create a single segmentation of each tissue (instead of thresholding each tissue separately).
  • https://github.com/gattia/DOSMA/blob/adcebe3df0556eaddac95b3d7501d313f361a367/dosma/models/stanford_qdess_bone.py#L199
• Added option connected_only to model.generate_mask this is because the bones produced a few spurious points in tests. This is on by default for all tissues.
  • https://github.com/gattia/DOSMA/blob/adcebe3df0556eaddac95b3d7501d313f361a367/dosma/models/stanford_qdess_bone.py#L217
• The model segments medial/lateral tissues separately.
  • To keep convention with other models, it then creates a "men" and "tc" tissue that just combines med/lat segments.
    • https://github.com/gattia/DOSMA/blob/adcebe3df0556eaddac95b3d7501d313f361a367/dosma/models/stanford_qdess_bone.py#L165-L176
  • Also saved the "all" segmentation mask that is a single mask with all tissues (not one-hot/binary).
    • https://github.com/gattia/DOSMA/blob/adcebe3df0556eaddac95b3d7501d313f361a367/dosma/models/stanford_qdess_bone.py#L152-L154

The main things I see myself wanting to do/add are:

• [ ] Create function to build original model so we can skip resampling images up/down?
• [ ] Add a model using same architecture as other stanford_qdess.py models - I think the segmentation performance on this model vs. that are pretty much the same.

[ad12]

@gattia thanks for the PR! could you lint your files and provide some same screenshots of the segmentations

[gattia]

Looks like it's getting hungup during CI because Python 3.6 isn't available anymore?

Performed lint on my side + added a photo of the current segs. I'd love to add the two remaining things I have listed on my PR comments, but they're not necessary for this PR and maybe I can do at a later date.