mrinaljain17 / ct-image-segmentation Goto Github PK

The transformations should be applied to both images and the masks (except like normalization)

• Windowing: Use brain, bone, and soft tissue windowing as 3 stacked channels. Added in #22
• Normalization: Get normalized values corresponding to each windowing type. Done in commit d7996cc
• Defaults: Implement some set of base/default transformations that can be used directly in a dataloader. Added a sample in commit 5538bc1. Other transformations can be experimented later, with a similar format.

For example: how are the pixels inside a "Mandible" are distributed?

This will be at two levels:

1. For individual patients, how does the pixel (or "voxel") value vary across different regions. Added in commit e0505dc
2. Across all patients, what's the distribution of these pixels for each region. *Added in commit 7dca68e

This could also be useful for creating very simple baseline models, essentially based on stuff like thresholding. Will also be helpful for windowing.

Ideally, should have the functionality to choose one or all the segmentation regions as targets.

This is required for working with 2D models, taking individual slices as inputs.

• Not completely sure, but I think they can be used to remove noise elements from the CT scans.

This could be useful for visualizing/evaluating model outputs later, when we don't have explicit paths to the 3D structures.

Is windowing needed?

• How will it help? - We might want to "highlight" certain regions for better segmentation.
• Possible add functionality to perform windowing to the existing Patient or Volume classes. A better way would be to use windowing as PyTorch transformations (with some sort of randomization).

Need more discussion!

For some reason, based on the default splitting (taken from actual competition), all the scans in the validation set are without mandible annotation. This can be fixed by simply resampling the examples from the combined train and validation sets, effectively creating new splits.

Note that the test data should not be used at any cost for this.

Does the transform we apply in the getitem to the images/masks include a collection of transformations already or do we need to add a loop if we want to have more than one transformation? (have not used Albumentations before)

This is definitely important for 3D models.

• Even for 2D models, if we want to crop unnecessary pixels (chest region, the CT scanner artifacts, etc), it might important to have all the scans on the same scale.

Code would be something along the lines of:

from scipy.ndimage import zoom

patient = ...

required_spacing = np.array([1., 1., 1.])
resize_factor = patient.image.spacing / required_spacing
vol = patient.image.as_numpy()[0]

resampled = zoom(vol, resize_factor, order=2, mode="nearest")

• Resampling could be really slow!

Each CT scan is of dimension (# slides, 512, 512), where a lot of volume is unnecessary - maybe due to scanner artifacts, or slices below the neck etc. Add functionality to crop only "useful" patch of volume from the entire CT scan.