GM segmentation performance on cropped images

[Nilser3]

Description

The seg_gm_contrast_agnostic (release: r20250204 ) was trained in multiple contrasts #2 (on full axial plane FOV) with very interesting results, but some non-optimal results have been observed in cropped images,
so in this issue we are going to analyze the performance of this model on cropped images in the SC mask, with different dilation factors.

Proposition

1. 10 subjects were selected from sct-testing-large (T2star contrast, 512 x 512 x 20 matrix, resolution: 0.3516 x 0.3516 x 3.3).
2. Automatic segmentation of SC with seg_sc_contrast_agnostic SCT v. 6.5, GM with sct_deepseg_gm SCT v. 6.5 and seg_gm_contrast_agnostic release: r20250204.
3. Cropping of the anatomical images and GM masks using -m SC_seg -dilate with following dilation factors:

-dilate 10	-dilate 20	-dilate 30	-dilate 40	-dilate 60	-dilate 90

-dilate 120	-dilate 150	-dilate 180	-dilate 210	-dilate 240	-dilate 270 (full FOV)

4. Re-segment the cropped anatomical image using sct_deepseg_gm SCT v. 6.5 and seg_gm_contrast_agnostic release: r20250204.
5. Calculate the Dice Score between the native cropped and re-segmented GM masks for each method.

Results

• sct_deepseg_gm is more robust to cropping than the seg_gm_contrast_agnostic model.
• seg_gm_contrast_agnostic model is impacted by cropping from 180 and lower dilation factors.

Related issues

#7

[Nilser3]

TODO:

• Re-train a model with modifications on:

transforms.append(
            SpatialTransform(
                patch_size_spatial, patch_center_dist_from_border=0, random_crop=False, p_elastic_deform=0,
                p_rotation=0.2,
                rotation=rotation_for_DA, p_scaling=0.2, scaling=(0.7, 1.4), p_synchronize_scaling_across_axes=1,
                bg_style_seg_sampling=False  # , mode_seg='nearest'
            )

see: nnUNetTrainer.py

[jcohenadad]

this is a very interesting comparison @Nilser3 👏 Note that sct_deepseg_gm was trained using a dilated convolution architecture-- maybe this is related?