Designing an Intelligent Lesion Detection System Using Deep Architecture Neural Networks in the Lower Limb X-Ray Images

Purpose: Diagnosis of musculoskeletal abnormalities is critical because of the large number of people affected by these disorders worldwide. The recent advances in deep learning techniques show that convolutional neural networks can be a useful tool for the computer-aided detection of radiographic abnormalities. This study focuses on diagnosing musculoskeletal abnormalities in the lower extremities using X-Ray images by deep architecture neural networks. Materials and Methods: The dataset contains 61,098 musculoskeletal radiographic images, including 42,658 normal and 18,440 abnormal images. Each image belongs to a single type of lower extremity radiography, including the toe, foot, ankle, leg, knee, femur, and hip joints, which were prepared with standard projection without artifacts and with high quality. A novel deep neural network architecture is proposed with two different scenarios that perform the lower extremity lesion diagnosis functions with high accuracy. The foundation of the proposed method is a deep learning framework based on the Mask Regional Convolutional Neural Network (R-CNN) and Convolutional Neural Network (CNN). The model with the best results incorporated the Mask R-CNN algorithm to produce the bounding box, followed by the CNN algorithm to detect the class based on that. Results: The proposed model can identify different types of lower limb lesions by an Area Under the Curve (AUC) of the Receiver Operating Characteristics (ROC) curve 0.925 with an operating point of 0.859 of sensitivity and a specificity of 0.893. Conclusion: The results indicated that the consecutive implementation of Mask R-CNN and CNN has a higher efficiency than Mask R-CNN and CNN separately in lesion detection of lower limb X-ray images.