XAI-CBM: An Explainable AI Framework with Concept Bottleneck Model for Automated Knee Osteoarthritis Severity Grading from X-ray Images

Abstract

Knee osteoarthritis (KOA) is a major cause of long-term musculoskeletal disability around the world, affecting more than 365 million people. For making clinical decisions, it is very important to accurately and quickly grade the severity of KOA using the Kellgren–Lawrence (KL) classification system. However, traditional radiographic assessment has a lot of variability between observers and is not very clear. This paper introduces XAI-CBM, an Explainable Artificial Intelligence framework that combines a ResNet-18 convolutional neural network (CNN) encoder with a dual-branch Concept Bottleneck Model (CBM) to automate and interpret KOA severity grading from standard knee X-ray images. The suggested architecture predicts KL grades (0–4) at the same time that it makes a concept vector that can be understood and includes ten clinically relevant radiographic proxies, such as joint-space width estimation, edge density, gradient magnitude, and intensity symmetry. There are three modules that work together to explain things: Gradient-weighted Class Activation Mapping (Grad-CAM) for showing spatial attention, SHAP (SHapley Additive exPlanations) for attributing features at the concept level, and gradient-based counterfactual concept optimization for giving clinical feedback that can be acted on. XAI-CBM was trained and tested on the publicly available SilpaCS/kneeosteoarthritis dataset, which has 3,000 images and stratified splits. It got 99.12% classification accuracy, 98.87% macro F1-score, and 99.41% macro ROC-AUC across five severity classes. A comparative analysis with eight cutting-edge methods shows that this one is always better. The framework is made for reliable collaboration between humans and AI in clinical radiology. It gives radiologists model reasoning that fits with established diagnostic standards.