Automatic hip fracture detection using deep learning

Abstract

Hip fractures represent a significant public health problem in Hong Kong and around the world. Automatic detection of hip fractures using deep learning on X-ray images leads to more accurate and efficient diagnosis and thus improves patient outcomes. In this study we developed an automatic hip fracture detection system which consists of a pipeline of four neural network models, namely the hip/pelvis model, the localization model, the prosthesis model and the fracture model, designed to take the patient X-rays from an examination session as input and output a hip fracture diagnosis in an end-to-end manner. Translated to clinical practice, the system can automatically identify hip and pelvic X-rays, locate the target fracture area, exclude irrelevant cases with prostheses, and output fracture diagnosis. We collected the hip and pelvic X-rays taken across 43 public hospitals and institutions in Hong Kong between 1 January 2008 and 31 December 2017, covering a 10-year time range and region-wide geological spread which is representative of the overall patient demographics of Hong Kong. We proposed a simple and efficient data cleaning paradigm to deal with large dataset with noisy labels and built a quality dataset of over 500,000 hip and pelvic images with clean labeling from over 180,000 patients. The trained fracture detection model presents diagnostic performance superior to human experts, and robustness towards data heterogeneity inherent in medical images. When compared against recently published work on automated hip fracture detection, our model surpasses the compared models by a large margin in all evaluation metrics. The completed system was assessed by and delivered to the Hospital Authorization of Hong Kong as one of the first medical AI systems to go into practical use and serve for the betterment of public health.

Date of Award	2020
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Kwang-Ting Tim CHENG (Supervisor)