Multi Locale Bone Fracture Radiographs and Localization

Abstract

We introduce MLBFR, a varied radiographs dataset of human bone fractures. The dataset contains 2,583 radiographs, among which 410 have 575 fracture points. A radiologist manually labelled the dataset as ”fractured” and ”non-fractured” with masks for the fracture locations. The dataset was verified and approved by an expert medical officer to evaluate the radiologist’s performance further. To precisely detect and localize the fracture areas, we experimented with several state- of-the-art object detection models, YOLOv5, maskRCNN, efficientDet and more, along with their ensemble. The trained models fell under two criteria, one being the full dataset and the other being only the fractured radiographs. The trained models managed to achieve a precision of 78.9% and 91.65% on combined and only fractured radiographs, respectively. The model performances were comparable to that of radiologists in detecting major abnormalities in the arm and shinbone area. With falling slightly behind in detecting fractures in the hip, thigh, and finger fractures. It is our belief that the task of improving this performance will be a good challenge for future research. To further encourage advancement in this area, we intend to make this dataset freely available in the future.