ThermoJointNet: A Hybrid CNN–DBN Framework for Early Detection of Knee Osteoarthritis using Thermal Imaging

Background: Osteoarthritis (OA) is a prevalent degenerative joint disorder characterized by the progressive breakdown of cartilage and inflammation, leading to pain and reduced mobility. Traditional diagnostic techniques, such as radiographic imaging, often fail to detect OA in its early stages, limiting timely intervention. In contrast, thermal imaging has emerged as a promising non-invasive alternative, capable of capturing temperature variations associated with inflammatory responses in the joints. Objective: This work proposes a novel approach that integrates a Modified Deep Belief Network (DBN) within a Convolutional Neural Network (CNN) architecture for early OA detection using thermal imaging. The primary goal is to enhance diagnostic accuracy, sensitivity, and specificity, thereby improving early-stage OA identification and classification. Methods: Thermal images from 1,000 participants were acquired under controlled environmental conditions. Images underwent multi-stage preprocessing involving normalization, bilateral filtering, and morphological enhancement. Regions of interest were segmented using K-means clustering. Dual feature extraction was performed using statistical descriptors and a convolutional neural network (CNN). These fused features were classified using a Modified Deep Belief Network (DBN) comprising stacked Restricted Boltzmann Machines with dropout regularization. Performance was evaluated using accuracy, sensitivity, specificity, and AUC. Results: The proposed method achieved a classification accuracy of 96%, with 95% sensitivity, 94% specificity, and an AUC of 0.95, surpassing existing state-of-the-art approaches. The system demonstrated the potential for reliable early-stage OA detection and classification based on thermal imaging. Conclusion: The proposed modified DBN based approach offers a highly accurate, non-invasive tool for OA diagnosis, paving the way for timely interventions and effective disease management. Future work will focus on integrating meta-heuristic optimization techniques to enhance the performance and expand the model's applicability to other medical imaging tasks.