Deep Learning Based Method To Identify And Detect Fractures From Musculoskeletal X-rays

Purpose: Identifying fractures on radiographs is a strenuous task due to the considerable variation of bone structures. Moreover, enormous patient load, vast disease classification, and lack of radiologists contribute to the misclassification of fractures and further delay in treatments. Automatic fracture detection techniques can assist in preventing the missing of unnoticeable fractures and enable the care provider to focus more on the treatment pathway. The aim of this work is to build a robust deep-learning system capable of accurately detecting fractures from radiographs. This system can assist clinicians in accurately and timely diagnosing fractures and further improving morbidity related to fractures and complications.

Methods/Materials: A total of 600,000 x-ray scans along with their clinical radiology reports were collected from multiple centers across India. The scans were collected and annotated by the radiologists in past 1 year. The deep neural network was trained on 450,000 and tested on 150,000 x-rays. We trained a single model classification and segmentation network which can detect and segment fractures in 16 different body parts. The fracture types include healed, chronic, hairline, greenstick, acute, lateral, oblique, frontal, displaced, and undisplaced fractures. The standard performance evaluation metrics are utilized to report the performance of the deep learning system.

Results: The deep learning system achieved an overall AUC of 0.95 (C.I. 0.945-0.9564), a sensitivity and specificity of 0.8956 and 0.9273 respectively, for fracture detection on all the body parts in 150,000 test data, with IoU of 0.65. The highest AUC was achieved in the femur, leg, forearm, finger, hand, pelvis, and toe with AUC greater than 0.93. In all 16 body parts, the IoU is greater than 0.55 and AUC is more than 0.90. Detailed analysis and performance metrics are provided in the supplementary material.

Conclusions: We present a deep learning-based automated reliable method to detect fractures fast and accurately. The achieved performance metrics are suggestive of the model's potential to identify the body parts and fractures and for further evaluation in real clinical settings.