license: mit
tags:
- medical-imaging
- knee-osteoarthritis
- kellgren-lawrence
- x-ray
- multimodal
- pytorch
KneeVision++ — Knee Osteoarthritis Severity Grading Checkpoints
Inference checkpoints from KneeVision++, an explainable multimodal pipeline for grading knee osteoarthritis severity (Kellgren-Lawrence 0–4) from X-rays and/or structured clinical/radiographic text.
Live demo: Darshan13/kneevision-demo
Files
| File | Model | Task | Test Quadratic Kappa |
|---|---|---|---|
best_densenet121.pt |
DenseNet121 | 5-class KL grading (image) | 0.775 |
best_efficientnet-b4.pt |
EfficientNet-B4 | 5-class KL grading (image) | 0.640 |
best_convnext_small_binary.pt |
ConvNeXt-Small | binary OA screening (image) | 0.757 |
best_densenet121_binary.pt |
DenseNet121 | binary OA screening (image) | 0.719 |
best_efficientnet-b4_binary.pt |
EfficientNet-B4 | binary OA screening (image) | 0.658 |
best_clinical.pt |
BioClinicalBERT | 5-class KL grading (clinical text) | 0.953 |
best_fusion.pt |
CNN + BioClinicalBERT fusion | 5-class KL grading (multimodal) | 0.959 |
Each best_*.pt has a matching best_*.json sidecar with {model_name, num_classes, binary, ordinal, image_size, best_kappa} metadata used by the loader.
Training data
- Images: Kaggle Knee Osteoarthritis Dataset — 8,260 X-rays, KL grades 0–4
- Clinical text: composed from real Osteoarthritis Initiative (OAI) fields — demographics, WOMAC pain/stiffness/function, and per-compartment OARSI radiographic grades (joint space narrowing, osteophytes, sclerosis, attrition) — not the KL label itself
Test-set results (1,656 held-out X-rays)
| Modality | Accuracy | Quadratic Kappa |
|---|---|---|
| Image only (DenseNet121) | 60.3% | 0.792 |
| Clinical Text only (BioClinicalBERT) | 87.4% | 0.953 |
| Multimodal Fusion | 88.9% | 0.959 |
Caveat: the clinical-text branch is fed real per-compartment OARSI fields that mechanically define the KL grade, not free-text patient-reported symptoms — so this is closer to "decode KL's own defining components written as prose" than "infer severity from symptoms alone." Fusion only marginally beats text-alone (+1.5pp accuracy) because the text branch already carries near-complete radiographic signal. Full discussion in the project README.
Usage
from huggingface_hub import hf_hub_download
import torch
path = hf_hub_download(repo_id="Darshan13/KneeVision-models", filename="best_densenet121.pt")
state = torch.load(path, map_location="cpu", weights_only=False)
See src/kneevision/models/image_model.py
(load_trained_model) for the full loading logic (backbone reconstruction, ordinal-head detection,
multitask auxiliary-head detection).
Disclaimer
Research/educational checkpoints only. Not a medical device, not validated for clinical use, not intended to inform real diagnostic or treatment decisions.