Abstract:
Musculoskeletal diseases require immediate attention to avoid chronic issues since they impact over 1.7 billion individuals globally. Radiographic images frequently identify musculoskeletal illnesses and abnormalities in medical computer vision. Researchers have suggested many methods for this purpose. In this research, we suggest five state-of-the-art transfer learning models with a partial layer-freezing technique where initial layers of a pre-trained model were frozen, except the last ten layers, to identify such Musculoskeletal Abnormalities using the MURA dataset, one of the most extensive collections of upper extremity radiographs. This method outperforms the performance of the baseline model in finger and humerus studies by achieving Cohen's kappascore of 0.439 in the DenseNet169 model and 0.638 in the DenseNet121 model. Although the current approach faces challenges in reaching sufficient accuracy for wrist, elbow, forearm, hand,and shoulder studies, there are positive improvements. When applied to radiographic images, the DenseNet network surpasses the ResNet, Inception, and Xception networks when evaluated using five different evaluation metrics. This shows that progress is being made in enhancing the performance of neural networks for improved medical image processing and analysis. Keywords: Binary Classification; Convolutional Neural Network; Transfer Learning; Layer- freezing; Dense Net network; ResNet; Inception; and Xception networks
