Automating Knee Osteoarthritis Severity Grading with CNN-Attention Hybrid Models

To this end, this research proposes a dual CNN attention model to self-attend and assign severity to KOA autonomously. It employs CNNs for feature extraction and attention processes in an effort to enhance the interpretability over personal judgments and variations during manual evaluation. The recommended pipeline using publically available datasets like OAI and MOST had an average accuracy of 92.0 percent in diagnosing the severity classes of KOA, thus justifying its reliability and effectiveness. Other key performance indicators also emphasize the model's stability to some extent. For all severity levels, the hybrid method achieved the assessment of average precision, recall, and F1-scores of 0.89, 0.88, and 0.89, fortunately. All the grades' ROC-AUC scores were above zero point nine, with the maximum of point nine five for the grade two indication of the model's capability of classifying different KOA severity levels with good levels of certainty. The Grad-CAM visualizations made it experientially evident that the model is focusing on features clinically relevant - the joint space narrowing, as well as the presence of osteophytes – further enhancing the model's clinical interpretability. The identical average accuracy for all the levels of severity and minimum misclassification errors that have been shown in the confusion matrix substantiate the practical usefulness of the model in clinical practice. The developed approach could potentially improve patient's and the general health care enterprise's benefit and effectiveness through the objectification of the diagnosis process. Future work will focus on increasing the variety of the dataset to enhance the utility and the potential usage of the additional modalities.