Abstract
BACKGROUND: Knee osteoarthritis (KOA) constitutes the prevailing manifestation of arthritis. Radiographs function as a common modality for primary screening; however, traditional X-ray evaluation of osteoarthritis confronts challenges such as reduced sensitivity, subjective interpretation, and heightened misdiagnosis rates. The objective of this investigation is to enhance the validation and optimization of accuracy and efficiency in KOA assessment by utilizing fusion deep learning techniques. METHODS: This study aims to develop a highly accurate and lightweight model for automatically predicting and classifying KOA through knee X-ray imaging. We propose a deep learning model named OA-MEN, which integrates a hybrid model combining ResNet and MobileNet feature extraction with multi-scale feature fusion. This approach ensures enhanced extraction of semantic information without losing the advantages of large feature maps provided by high image resolution in lower layers of the network. This effectively expands the model's receptive field and strengthens its understanding capability. Additionally, we conducted unseen-data tests and compared our model with widely used baseline models to highlight its superiority over conventional approaches. RESULTS: The OA-MEN model demonstrated exceptional performance in tests. In the unseen-data test, our model achieved an average accuracy (ACC) of 84.88% and an Area Under the Curve (AUC) of 89.11%, marking improvements over the best-performing baseline models. These results showcase its improved capability in predicting KOA from X-ray images, making it a promising tool for assisting radiologists in diagnosis and treatment selection in clinical settings. CONCLUSION: Leveraging deep learning for osteoarthritis classification guarantees heightened efficiency and accuracy. The future goal is to seamlessly integrate deep learning and advanced computational techniques with the expertise of medical professionals.