sash99666/cdti-lv-baseline

cDTI LV segmentation — baseline weights (MD + DWI)

Weights for the automated whole-heart cardiac DTI pipeline: YOLO heart crop → resize 256 → nnUNet ResEnc-M LV segmentation (5-fold) → YOLO RV insertion points → un-crop to original FoV. Easiest way to run it: the one-command tool in the deploy/ folder of the GitHub repository sash9966/nnUNet_cDTI (branch master) — its README walks through install, download and input format. The research notebook (PaperRun_cDTI_Pipeline.ipynb) is on branch crop-alignment-and-hap-fixes.

Contents

nnUNet_results/Dataset310_HannumSmartHealthandDirVsAvgsLVMDandAvg/
  nnUNetTrainerDA5_100epochsNoCompile__nnUNetResEncUNetMPlans__2d/
    dataset.json  plans.json  dataset_fingerprint.json
    fold_0/ ... fold_4/checkpoint_final.pth      # 5-fold ensemble, ~848 MB each
yolo/
  crop.pt                                       # heart-crop detector (average DWI)
  rvip.pt                                       # RV insertion-point detector (average DWI)

Keep the folder names exactly as they are — nnUNet locates the model by them.

Inputs (per slice, after the YOLO crop + resize to 256×256)

channel	image	normalization
_0000	average diffusion-weighted image (DWI)	min–max to [0, 1]
_0001	mean diffusivity (MD)	divide by 4.0

nnUNet normalization is off (noNorm in dataset.json), so apply exactly this normalization yourself — a mismatch silently degrades the segmentation.

Usage

hf download sash99666/cdti-lv-baseline --local-dir cdti_baseline
export nnUNet_results=$PWD/cdti_baseline/nnUNet_results
export nnUNet_compile=false
nnUNetv2_predict -i <imagesTs_256> -o <out> -d 310 -c 2d \
  -tr nnUNetTrainerDA5_100epochsNoCompile -p nnUNetResEncUNetMPlans -f 0 1 2 3 4

In the notebook, set COLLAB['lv_results_dir'] to cdti_baseline/nnUNet_results and COLLAB['crop_weights'] / COLLAB['rvip_weights'] to cdti_baseline/yolo/crop.pt / rvip.pt.

Performance (held-out test set: 39 slices, 6 volunteers, never seen in training)

metric	value
LV Dice (mean)	0.909
Precision / Recall	0.889 / 0.932
MD / FA / DWI median agreement	~1% / ~2% / ~3%
Insertion-point Hausdorff (median)	~2.3–2.4 mm (assumes 1.25 mm pixels)
Helix angle pitch, manual vs automatic mask (mean)	−61.0 vs −61.7 °/wall (median per-slice difference 4.2)

Limitations

• Single center, single protocol (DiVO_15_04); other scanners/protocols are unvalidated.
• Expects the average DWI and MD maps already reconstructed (no tensor fitting).
• The insertion-point detector is weak (a few slices miss by 30–55 mm); treat its points as an initialization.
• .pth checkpoints are loaded with torch.load; only use files from this repository.