Skip to content

Patch_Sampling.md

Latest commit

 

History

History
39 lines (31 loc) · 4.27 KB

Patch_Sampling.md

File metadata and controls

39 lines (31 loc) · 4.27 KB
Raw

Patch Sampling during training

The neural network is trained on 3D patches sampled from the larger 3D MRI volume. For example, the MRI volume may be of shape (1, 44, 638, 638), but the neural network is trained on patches of size (1,20,256,256).

One of the training strategies that has proven effective has been to overample the patches centered on the ground truth segmentation region, rather than randomly sampling patches from anywhere in the input volume. This patch sampling strategy ensures there are non-zero segmentation values within the volume. Sampled patches can be centered on points from any of the 3 segmentation regions -- atrium wall, left atrum cavity, or right atrium cavity.

A (1,20,256,256) sampled patch is quite large. Realistically, most sampled patches will contain most of the heart volume. image

Patch sampling that ensures full coverage of the depth (z) dimension

By default, the oversampling patch sampler will randomly sample patches centered at one of the predefined sampling points (the dots in the above figure). There area of the heart region is smaller for slices towards the edge of the 3d volume. For example, for a volume of shape (1,44,638,638), the area of the visible shape in slice 5 is significantly smaller than the area in slice 25. Therefore, the random patch sampler is less likely to sample a patch centered on a slice with an index towards the edge of the 3d volume.

I introduced a patch sampling policy to ensure that the depth (z) dimension is equally sampled by enforcing all slices are sampled.

model Rank Avg_Rank DSC_wall HD95_wall DSC_right HD95_right DSC_left HD95_left
8 nnUNetTrainer__nnUNetResEncUNetLPlans__3d_fullres2 9 14 0.725331 2.76333 0.92567 3.20032 0.930359 3.68754
9 mbasTrainer__plans_2024_08_30__ResEncUNet_p20_256_dil2_batch_dice_cascade_ResEncUNet_08_27 10 15.3333 0.723445 2.72685 0.925043 3.21474 0.931448 3.71448
83 mbasTrainer__plans_2024_09_04__ResEncUNet_p20_256_dil2_bd_zcov_cascade_ResEncUNet_08_27 12 16.6667 0.722161 2.74292 0.925147 3.24111 0.931686 3.69314
85 mbasTrainer__plans_2024_09_04__ResEncUNet_p16_192_dil2_bd_zcov_cascade_ResEncUNet_08_27 20 23 0.721729 2.77263 0.924553 3.28032 0.931161 3.8619
27 mbasTrainer__plans_2024_09_02__MedNeXtV2_p16_256_dil2_nblocks346_slim128_cascade_ResEncUNet_08_27 29 32.5 0.718305 2.80845 0.921939 3.31858 0.929709 4.06074
33 mbasTrainer__plans_2024_09_04__MedNeXtV2_p16_256_dil2_bd_zcov_nblocks346_slim128_cascade_ResEncUNet_08_27 35 38.6667 0.71979 2.97324 0.920986 3.4197 0.929486 4.07476

Comparing ResEncUNet_p20_256_dil2_batch_dice_cascade_ResEncUNet_08_27 vs. ResEncUNet_p20_256_dil2_bd_zcov_cascade_ResEncUNet_08_27`.

  • training standard sampling without forcing the full depth (z) dimension coverage appears to be better. Comparing ResEncUNet_p20_256_dil2_bd_zcov_cascade_ResEncUNet_08_27 vs. ResEncUNet_p16_192_dil2_bd_zcov_cascade_ResEncUNet_08_27
  • I hypothesized that the smaller patch size would make it less likely for the model to overfit since there's increased diversity between small patches.
  • Models trained on the samller patches performed worse. Comparing MedNeXtV2_p16_256_dil2_nblocks346_slim128_cascade_ResEncUNet_08_27 vs. MedNeXtV2_p16_256_dil2_bd_zcov_nblocks346_slim128_cascade_ResEncUNet_08_27
  • Again, the standard sampling strategy performed better than the new z-dimension coverage sampling.

Unfortunately the new sampling policy doesn't seem to have improved performance.