Hybrid Deep Learning Model for Classification and Prediction of Abnormalities in Upper and Lower Extremities of Musculoskeletal Radiographs

Musculoskeletal disorders are a significant global public health concern, comprising various conditions affecting the skeletal system and its associated structures. These disorders present challenges in terms of diagnosis, treatment, and management due to their complex nature, diverse clinical manifestations, and multifactorial causes involving biological, environmental, and socioeconomic factors. Traditional methods in the diagnosis of musculoskeletal disorder rely significantly on the expertise of the radiologists, which can be affected by various factors such as fatigue, experience level and subtlety of some abnormalities. Moreover, these methods are normally time-intensive, highlighting the need for more efficient and consistent diagnostic approach. To date, no model exists that can predict abnormalities in both upper and lower extremities. In this study, we proposed a novel deep learning model designed to comprehensively cover the entire musculoskeletal system. The model classifies input radiographs into one of eleven classes (elbow, finger, shoulder, wrist, hip, humerus, ankle, foot, knee, forearm, and hand) and predicts the presence of abnormalities. Our model operates in a pipeline fashion: the input image first undergoes preprocessing, where features are extracted using three specialized modules. The processed image is then passed through convolutional and Squeeze-and-Excitation blocks for training. Finally, the model classifies the input radiographs into one of the eleven categories and predicts whether the radiograph is normal or abnormal. The proposed model achieves outstanding results, first classifying input radiographs into one of eleven study types with an accuracy of 97.37%, and subsequently identifying each radiograph as either normal or abnormal with an accuracy of 89%. Additionally, the model achieves a ROC-AUC of 0.94, a sensitivity of 0.86, a specificity of 0.89 and Cohen’s kappa of 0.77. Overall, these result highlights the superior performance of our model in both classifying X-ray images into correct categories and accurately predicting the presence of abnormalities, surpassing all existing models in these aspects. This innovative approach has the potential to enhance clinical decision-making, facilitate early detection of abnormalities, and ultimately improve patient outcomes in musculoskeletal health. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2024.