Fracture Classification in Musculoskeletal Radiographs Using Custom CNN and Ensemble Learning

Musculoskeletal traumas, specifically fractures, pose significant hurdles for healthcare systems on a global scale. The conventional method of categorizing fractures heavily depends on the proficiency of radiologists, which introduces the possibility of mistakes and impedes precise diagnosis. By utilizing digital radiography, our objective is to mitigate the constraints associated with conventional approaches and boost the effectiveness and dependability of fracture categorization. Our investigation expands on this framework by presenting a customized Convolutional Neural Network specifically engineered for musculoskeletal radiographic images. To further augment classification precision and resilience, we integrate adapted pre-trained models with tailored layers as well as Ensemble Learning, amalgamating the capabilities of several models. The fusion methodology endeavors to alleviate hurdles pertaining to data scarcity, providing a robust framework for enhancing automated fracture detection systems in healthcare environments. Expanding upon recent efforts in transfer learning for fracture detection, our proposed approach seamlessly integrates into current research. By combining a customized Convolutional Neural Network (CNN) with Ensemble Learning, we introduce a resilient framework primed to enhance automated fracture identification systems. Our results strongly support the incorporation of adapted DenseNet121 with tailor-made layers, outperforming all alternative models by achieving a remarkable accuracy of 93%. This advancement represents a significant breakthrough in the enhancement of fracture and musculoskeletal injury diagnosis and treatment. This will also facilitate radiologists and physicians in expediently discerning fractures, enabling a more targeted approach to treatment and reducing the timeframe required to identify and pinpoint the specific locations of the fractures. Due to lightweight characteristics of the model, portable handheld instruments can be utilized for identification purposes with ease.