Hybrid Deep Learning Approach for Automatic Detection in Musculoskeletal Radiographs

Simple Summary Musculoskeletal disorder is affecting a large population globally and is becoming one of the foremost health concerns. The treatment is often expensive and may cause a severe problem in the case of misdiagnosis. Therefore, a reliable, fast, and inexpensive automatic recognition system is required that can detect and diagnose abnormalities from radiographs to support effective and efficient decision making for further treatment. In this work, the finger study type from the MURA dataset is taken into consideration owing to the fact the existing models were not able to give the desired performance and accuracy in detecting abnormalities in finger radiographs. Herein, a novel deep learning model is proposed, wherein after the preprocessing and augmentation of the finger images, they are fed into a model that learns the discriminative features through multiple hidden layers of dense neural networks and classifies them as normal or abnormal radiographs. The achieved result outperforms all existing state-of-the-art models, making it suitable for clinical settings. This will help society in the early detection of the disorder, which reduces the burden on radiologists and reduces its long-term impact on a large population. The practice of Deep Convolution neural networks in the field of medicine has congregated immense success and significance in present situations. Previously, researchers have developed numerous models for detecting abnormalities in musculoskeletal radiographs of upper extremities, but did not succeed in achieving respectable accuracy in the case of finger radiographs. A novel deep neural network-based hybrid architecture named ComDNet-512 is proposed in this paper to efficiently detect the bone abnormalities in the musculoskeletal radiograph of a patient. ComDNet-512 comprises a three-phase pipeline structure: compression, training of the dense neural network, and progressive resizing. The ComDNet-512 hybrid model is trained with finger radiographs samples to make a binary prediction, i.e., normal or abnormal bones. The proposed model showed phenomenon outcomes when cross-validated on the testing samples of arthritis patients and gives many superior results when compared with state-of-the-art practices. The model is able to achieve an area under the ROC curve (AUC) equal to 0.894 (sensitivity = 0.941 and specificity = 0.847). The Precision, Recall, F1 Score, and Kappa values, recorded as 0.86, 0.94, 0.89, and 0.78, respectively, are better than any of the previous models'. With an increasing appearance of enormous cases of musculoskeletal conditions in people, deep learning-based computational solutions can play a big role in performing automated detections in the future.