Spinal Implant Classifier — weights
ConvNeXt V2 ensemble (Nano/Tiny/Base) for cervical & thoracolumbar spinal-implant identification on radiographs. 30 fold/seed weights per size, fp16 safetensors. Research use only — not a medical device. See the demo Space for usage and the paper for methods and metrics. MIT license (same as upstream ConvNeXt V2).
Live demo: https://huggingface.co/spaces/NicholasJYee/spine-implant-classifier-demo
Archived (Zenodo): https://doi.org/10.5281/zenodo.20691495 (all versions)
Layout: weights/<cv2n|cv2t|cv2b>/fold<0-2>_seed<0-9>.safetensors
Implant classes (18)
Cervical (5): Medtronic Venture, Medtronic Zevo, Stryker Reflex Hybrid, DePuy Synthes Uniplate, Medtronic Atlantis.
Thoracolumbar (13): Medtronic Legacy, Precision Spine Reform, DePuy Synthes Click'X, DePuy Synthes EXPEDIUM, DePuy Synthes Matrix, Globus Medical CREO, Medtronic Solera, Alphatec Illico, Arcos Biomedical M8, DePuy Synthes Viper, Medtronic TSRH, Stryker Xia, Orthofix Firebird.
Usage
This repo ships a self-contained inference script (infer.py) with the model
definition (model_convnextv2.py) and label metadata (class_info.json).
pip install torch safetensors huggingface_hub pillow numpy
# fetch the helper files from this repo, then:
python infer.py xray.jpg --size cv2n --bbox 120 80 200 260
It downloads the chosen size's 30 weights, averages their logits, applies
temperature scaling, and prints the ranked implant classes. Pass a tight
bounding box around the implant for best results (the model is trained on
implant-centered crops); omit --bbox to use the whole image.