Using deep learning for abnormality detection in musculoskeletal radiographs

Abstract

Abstract One of the most remarkable applications of deep learning is in medical diagnoses and new improvements in this field have shown that with large enough datasets and right methods, one can achieve results as reliable as experienced doctors. One of such developments is MURA which is a dataset about musculoskeletal radiographs consisting of 14,863 studies from 12,173 patients, resulting in a number of 40,561 multi-view radiographic images. Each one of these studies is about one of seven standard upper extremity radiographic study types, namely, finger, forearm, elbow, hand, shoulder, homeruns, and wrist. Each study was categorized as normal or abnormal by board-certified radiologists in the diagnostic radiology environment between 2001 and 2012. Abnormality detection in muscular radiography is of great clinical applications. This gains more importance in cases which abnormality detection is difficult for physicians. If the proposed model can help us in detection, the process of treatment will precipitate. This model is termed inception-v3. In this study, we evaluated the MURA data set through Dense NET and inception-v3 methods. The results indicated that the former had better performance, and we added preprocessing module to it in order to improve the accuracy of the DenseNet method to detect abnormality. In this context, we train machine to be sensitive to presence of external objects to be distinguished with actual abnormality like bone fraction; we achieved that by a lot of various radiographs as machine inputs. By this strategy, both techniques (DenseNET and Inception-V3) showed improvement in accuracy. Thus we subgroup abnormality into with or without the presence of external objects. Although the average opinion of radiologists still shows better results, in images in which fracture detection is delicate, like finger fracture, the proposed model works more accurately, and it can as a decision support assistant for physicians in final detection of fracture. If the image is separated from normal images using Platinum, and a new class is made, and pre-processing is done, the precision of the proposed model enhances. So, a model which can automatically detect abnormality can identify the part of image which is detected to be abnormal by the model. If this model is efficient, it can interpret the images more efficiently, it can reduce errors, and it can enhance quality. In order to evaluate the integration of this model with other models of deep learning in clinical setting, more studies are needed to be carried out.