A Comparative Analysis of Deep Learning Models for Knee Osteoarthritis Severity Grading

Abstract

The Kellgren-Lawrence (KL) grading system is commonly used to evaluate knee osteoarthritis (OA), but it can be subjective and subject to variation among assessors. Our study looked at three Convolutional Neural Network (CNN) methods for OA severity classification from a dataset of 15,770 X-ray images to overcome this difficulty and create a more objective technique. Under the same preprocessing conditions, we contrasted a baseline custom CNN, DenseNet201, and a hybrid model with a CBAM attention mechanism. With an overall accuracy of 65%, a weighted precision and recall of 65%, and an F1-score of 64%, the hybrid model, which uses DenseNet201 as a fixed feature extractor, performed the best. This was better than both the baseline model (59% accuracy) and the standalone DenseNet201 (59% accuracy). Although the hybrid architecture has a lot of promise, we also had to deal with issues like overfitting. Our thorough comparison demonstrates how this hybrid strategy can successfully combine strong pre-trained features with the flexibility required for particular tasks. Although more clinical validation is necessary, this shows that automated systems like ours could improve diagnostic consistency in OA grading.   Keywords - Knee Osteoarthritis, Kellgren-Lawrence Grading, Deep Learning, Attention Mechanism, CBAM