Automatic bone fracture detection in x-ray images using deep learning
Abstract
Bone fractures are a leading cause of morbidity and mortality worldwide.
In Uganda, statistics for the prevalence of bone fractures are unknown, although
anecdotal evidence points to a high incidence, mostly arising from
traffic accidents and falls. The situation becomes worse year on year due to a
rising life expectancy, and thus an increasing number of the ageing population
who are more prone to fractures. To reduce the debilitating effects of these
fractures and improve quality of life, it is important that the fractures are
accurately diagnosed early on. X-ray imaging is the most common imaging
modality for fracture diagnosis in Uganda, but its manual interpretation is
usually error-prone, potentially leading to missed diagnoses. To address this
challenge, this project aimed at developing an automated bone fracture detection
system for efficient diagnosis, utilizing a deep learning approach.
Images of fractured bones were obtained from Roboflow and the open-source
dataset. We developed a model for the localization of bone fractures, utilizing
the YOLOv5 architecture. Our best model achieved a mean average precision
of 85.6%. Comparison with alternative approaches such as EfficientDet, and
Detectron2 reveals the superior performance of our model. Our model, when
integrated into a clinical decision support system, is potentially a promising
approach to improve clinical outcomes based on accurate and efficient bone
fracture detection from x-ray data.