Cross-Institutional Five-Class Kellgren-Lawrence Grading of Knee Osteoarthritis via Multitask Deep Learning

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models for Kellgren-Lawrence (KL) grading often exhibit optimistic performance due to data leakage.
• These models frequently fail to generalize across different institutions owing to domain shift, creating a reproducibility crisis.

Purpose of the Study:

• To introduce KL-FuseNet, a multitask deep learning architecture designed to fuse global and local features for improved KL grading.
• To address the challenges of data leakage and domain shift in automated osteoarthritis assessment.

Main Methods:

• Developed KL-FuseNet, a multitask architecture combining ConvNeXt-Base and ResNet-50 features for predicting ordinal grades, label distributions, and binary severity (KL≥2).
• Utilized strict patient-wise stratified splits on an internal Osteoarthritis Initiative dataset (n=8260) and an independent Chinese cohort (n=2295).
• Compared zero-shot transfer performance against a selective fine-tuning protocol for cross-institutional generalization.

Main Results:

• KL-FuseNet demonstrated robust internal agreement (QWK: 0.881, accuracy: 70.3%).
• External zero-shot deployment showed a domain gap (accuracy: 66.1%), but selective fine-tuning significantly improved performance (accuracy: 80.0%, QWK: 0.950, AUC for KL≥2: 0.984).

Conclusions:

• KL-FuseNet achieves state-of-the-art performance under rigorous evaluation, but domain-aware adaptation is crucial for clinical utility.
• The study establishes a reproducible method for deploying automated grading models across diverse medical centers, enhancing osteoarthritis diagnosis.