ADPKD L3 body composition — trained weights
Trained weights for Fully Automated Deep Learning for Third Lumbar-Level Body Composition Analysis on Abdominal Computed Tomography in Autosomal Dominant Polycystic Kidney Disease (Clinical Kidney Journal, 2026).
Code: https://github.com/OneChampOnly/adpkd-l3-body-composition
Given a 3D abdominal CT volume, the pipeline returns cross-sectional areas of skeletal muscle, visceral fat, subcutaneous fat and intramuscular adipose tissue at the L3 inferior endplate (primary) and L3 middle (secondary) reference sections.
CT volume ──▶ L3 vertebra segmentation ──▶ reference-section detection ──▶ body composition
3D nnU-Net (shared) YOLOv8n (one per level) 2D nnU-Net (one per level)
Contents
25 checkpoints, 4.66 GiB — five cross-validation folds × five models.
| Model | Framework | Path |
|---|---|---|
| L3 vertebra | nnU-Net v2, 3d_fullres |
nnUNet_results/Dataset920_L3Vertebra0917/…/fold_N/checkpoint_best.pth |
| Body composition, inferior | nnU-Net v2, 2d |
nnUNet_results/Dataset917_InferiorSMA300Reann0917/…/fold_N/checkpoint_best.pth |
| Body composition, middle | nnU-Net v2, 2d |
nnUNet_results/Dataset918_MiddleSMA300Reann0917/…/fold_N/checkpoint_best.pth |
| Detector, inferior | Ultralytics YOLOv8n | yolo/inferior/fold_N/best.pt |
| Detector, middle | Ultralytics YOLOv8n | yolo/middle/fold_N/best.pt |
Each nnU-Net model directory also holds plans.json, dataset.json and
dataset_fingerprint.json; nnU-Net needs them at inference time. SHA256SUMS covers every
file. Every number in the article comes from checkpoint_best.pth.
Usage
git clone https://github.com/OneChampOnly/adpkd-l3-body-composition.git
cd adpkd-l3-body-composition
python tools/download_weights.py --source OneChampOnly/adpkd-l3-body-composition --dest weights/
python tools/download_weights.py --dest weights/ --verify
python tools/download_weights.py --dest weights/ --write-registry config/model_registry.json
Then follow the repository README. For new scans any fold works; fold matching only matters when reproducing the article on the original cohort.
Training data and intended use
Trained on abdominal CT of patients with ADPKD from Korean institutions, with manual
annotation. The images and masks are not released (IRB and privacy restrictions); see
docs/DATA.md in the code repository. These weights are provided for research use and are
not a medical device.
Authors
Jigang Lee, DongGeun Yun, Yaerim Kim, Kyoung Tae Kim, Yong Chul Kim, Dha Woon Im, Hayne Cho Park, Youngwoo Kim
Citation
Please cite the article and, for the weights, the archived record:
Lee J, Yun D, Kim Y, Kim KT, Kim YC, Im DW, Park HC, Kim Y. Trained weights: fully automated L3-level body composition analysis on abdominal CT in ADPKD (v1.0.0). Zenodo; 2026. https://doi.org/10.5281/zenodo.22976872
See CITATION.cff in the code repository for full author metadata.