Classification of X-Ray Images for the Automated Severity Grading of Knee Osteoarthritis by Ensemble Learning Through EfficientNet Architectures with Grad-CAM Visualization

Knee osteoarthritis (KOA) is a degenerative form of arthritis in the knee joints commonly occurring in old-aged individuals which can cause extreme pain, deterioration of bone joints, joint membranes or ligaments, and irregularity in the knee joints. Being affected by KOA, individuals experience limited social interaction and diminished quality of life. The KOA can be assessed using the Kellgren-Lawrence (KL) grading method. In this study, an ensemble model was developed to predict a KL grading reliably and precisely. Several EfficientNet architectures were implemented for the ensembling of this study. Among them, an ensemble of EfficientNetB0 and EfficientNetB4 pre-trained models gave satisfactory results compared to other deep learning or ensembling models. Here, the proposed model underwent evaluation utilizing the Osteoarthritis Initiative (OAI) dataset which contains X-ray radiographs. The dataset was split into training, testing, and validation sets with percentages of 75%, 17%, and 8% respectively. The overall multiclass classification accuracy was 96.03% with an F1-score of 93%, specificity of 98.56%, and recall of 93%. The proposed model achieved an optimistic result compared to the transfer learning models and state-of-the-art models. Gradient-Weighted Class Activation Mapping was also applied for the radiographs to implement Explainable AI for generating visual explanations for the predictions made by the deep learning networks which revealed the model's potential for grading the severity of KOA.