Comparative Evaluation of Deep Learning and Foundation Model Embeddings for Osteoarthritis Feature Classification in...

Foundation models (FM) offer a promising alternative to supervised deep learning (DL) by enabling greater flexibility and generalizability without relying on large, labeled datasets. This study investigates the performance of supervised DL models and pre-trained FM embeddings in classifying radiographic features related to knee osteoarthritis. We analyzed 44,985 knee radiographs from the Osteoarthritis Initiative dataset. Two convolutional neural network models (ResNet18 and ConvNeXt-Small) were trained to classify osteophytes, joint space narrowing, subchondral sclerosis, and Kellgren-Lawrence grades (KLG). These models were compared against two FM: BiomedCLIP, a multimodal vision-language model pre-trained on diverse medical images and text, and RAD-DINO vision transformer model pre-trained exclusively on chest radiographs. We extracted image embeddings from both FMs and used XGBoost classifiers to perform downstream classification. Performance was assessed using a comprehensive classification metrics appropriate for binary and multi-class classification tasks. DL models outperformed FM-based approaches across all tasks. ConvNeXt achieved the highest performance in predicting KLG, with a weighted Cohen's kappa of 0.880 and higher AUC in binary tasks. BiomedCLIP and RAD-DINO performed similarly, and BiomedCLIP's prior exposure to knee radiographs during pretraining led to only slight improvements. Zero-shot classification using BiomedCLIP correctly identified 91.14% of knee radiographs, with most failures associated with low image quality. Grad-CAM visualizations revealed DL models, particularly ConvNeXt, reliably focused on clinically relevant regions. While FMs offer promising utility in auxiliary imaging tasks, supervised DL remains superior for fine-grained radiographic feature classification in domains with limited pretraining representation, such as musculoskeletal imaging.