Trained weights: fully automated L3-level body composition analysis on abdominal CT in ADPKD (v1.0.0)
Year: 2026
Type: software
Abstract: Trained weights for Fully Automated Deep Learning for Third Lumbar-Level Body Composition Analysis on Abdominal Computed Tomography in Autosomal Dominant Polycystic Kidney Disease. Version 1.1.0. The L3 vertebra segmenter is now nnU-Net v2 3d_lowres (v1.0.0 used 3d_fullres): same data, folds, trainer and checkpoint rule, but a patch covering about 288 × 243 × 243 mm instead of 168 × 131 × 109 mm. In the 300-scan out-of-fold evaluation, level misidentification fell from 4 to 2 scans at the inferior level and from 5 to 1 at the middle level. The detectors and body composition segmenters are byte-identical to v1.0.0. 25 checkpoints (4.66 GiB): five cross-validation folds × five models — a 3D nnU-Net v2 L3 vertebra segmenter (3d_lowres), YOLOv8n reference-section detectors for the L3 inferior endplate and L3 middle levels, and 2D nnU-Net v2 body composition segmenters (skeletal muscle, visceral fat, subcutaneous fat, intramuscular adipose tissue) for each level. Packaged in nnU-Net's nnUNet_results layout with plans.json, dataset.json and dataset_fingerprint.json. Files: adpkd-l3-weights-v1.1.0-fold0.tar.gz … -fold4.tar.gz (about 890 MB each; each holds one fold's five checkpoints plus the nnU-Net sidecar JSON files, so a single fold is usable on its own), SHA256SUMS for every file inside them, and ARCHIVES.sha256 for the archives themselves. Unpack the archives into weights/ of the code repository. Verify with python tools/download_weights.py --dest weights/ --verify. Code: https://github.com/OneChampOnly/adpkd-l3-body-compositionMirror: https://huggingface.co/OneChampOnly/adpkd-l3-body-composition No patient images or identifiers are included. For research use; not a medical device.
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223389752