code/whole_case_compare_patch_v2/README.md

CBCT Whole-Case Compare V2

This version fixes the overlap bug by baking every method-column translation directly into mesh vertices with mesh.apply_translation(...).

Final layout in ONE GLB:

LEFT -> RIGHT

1. Ground Truth
2. GeoCanal-SDF (Ours)
3. U-Mamba2
4. nnU-Net ResEnc L
5. MedNeXt-M k5

Install

cd /root/autodl-tmp
rm -rf CBCT_whole_case_compare_patch_v2
unzip -q CBCT_whole_case_compare_patch_v2.zip -d /root/autodl-tmp

bash /root/autodl-tmp/CBCT_whole_case_compare_patch_v2/install_to_cbct.sh

First test: Case033

cd /root/autodl-tmp/cbct
source /root/autodl-tmp/toothcanal_private/envs/core/bin/activate

rm -rf /root/autodl-tmp/toothcanal_private/clinical_whole_compare_v2/C033

python -m clinical_compare.whole_case_compare \
  --config benchmark_configs/private29_4fold_final.yaml \
  --cases 33 \
  --methods gt dualsdf_v21 umamba2 nnunet_resenc mednext \
  --gap-mm 35 \
  --preview \
  --strict \
  --out-dir /root/autodl-tmp/toothcanal_private/clinical_whole_compare_v2

Expected:

C033/
  C033_GT_Ours_SOTA_5dentitions.glb
  C033_GT_Ours_SOTA_5dentitions_canals_only.glb
  C033_preview_5dentitions.png
  layout.csv
  READ_ME_FIRST.txt

The terminal prints five separate X intervals. They must NOT overlap.

Final held-out cases

python -m clinical_compare.whole_case_compare \
  --config benchmark_configs/private29_4fold_final.yaml \
  --cases 31 32 33 34 35 \
  --methods gt dualsdf_v21 umamba2 nnunet_resenc mednext \
  --gap-mm 35 \
  --preview \
  --strict \
  --out-dir /root/autodl-tmp/toothcanal_private/clinical_whole_compare_v2

Spacing

Recommended: --gap-mm 35

If visually crowded: --gap-mm 50

If too far: --gap-mm 25

The gap is edge-to-edge empty space between complete dentitions.