Classification of Knee X-Ray Images by Severity of Osteoarthritis Using Skip Connection Based ResNet101

Abstract

Knee Osteoarthritis (KOA) severity currently de- pends on the Kellgren-Lawrence (KL) scale assessment, a method known for its subjective nature, leading to frequent disagreements among clinicians. To address this problem, we developed an automated, multi-task deep learning system. Our framework uses a ResNet101 architecture with skip connections, which perform two jobs at once: a detailed severity classification and a simple healthy-versus-osteoarthritic identification. We took a subset of the Osteoarthritis Initiative (OAI), which consists of 1,650 knee X-rays, and addressed the uneven class distribution by implementing techniques like focal loss and MixUp augmentation. We compared our model's performance against seven other CNN models (including VGG16 and UNet). Our multi-task ResNet101 outperformed all other models, achieving 87.10 % accuracy for detailed severity grading and 96.77 % accuracy for binary differentiation on unseen test data. This shows that residual networks with multi-task objectives and advanced regularization are a dependable and effective solution for automating KOA assessment.

Ähnliche Arbeiten