Automated classification of musculoskeletal abnormalities using a dynamic ensemble of deep vision models

Detecting and classifying musculoskeletal disorders is vital to preventing long-term disability and limited mobility; however, this procedure is challenging due to the shortage of radiologists. Existing artificial intelligence technologies, particularly deep vision models, aim to support specialists in rapid, accurate diagnosis. In this study, a dynamic ensemble model is developed using a Probability-based Dynamic Ensemble (PDE) approach with two vision models, ConvNeXt Base and Swin Transformer. Unlike conventional ensemble methods that employ fixed averaging or voting strategies, the proposed PDE framework learns both the contributions of individual backbone models and the optimal decision threshold directly from the training data, thereby exploiting the strengths of backbone models. Regional and image-based experiments are performed to detect abnormality in X-ray images of patients for seven anatomical regions of the MURA dataset. The radiographs are preprocessed and enhanced, and the models are trained individually. Dynamically determined probabilities of each model are used to construct and train an ensemble model with learnable weights and a threshold to mitigate the limitations of the backbone models. The proposed model obtained superior results in the majority of region-based experiments, with accuracies ranging from 0.826 to 0.920, and 0.843 in image-based experiments. Comparative statistical analyses are performed, and explainable AI tools are implemented to analyze clinical deployment. The obtained results demonstrate that the dynamically determined probabilities enable the ensemble model to optimize the balance between sensitivity and specificity by enhancing the individual recognition abilities of pre-trained models.