AI Tool for Hip Fracture Detection using Deep Learning and Computer Vision

A deep learning model to automate the process of identifying potential fractures by inspecting X-rays using Computer Vision for Tan Tock Seng Hospital

Background
Hip fractures are a leading cause of death and disability among older adults. Hip fractures are also the most commonly missed diagnosis on pelvic radiographs, and delayed diagnosis leads to higher cost and worse outcomes for the patient.

User Research and Limitations
There exists many limitations including multiple types of fractures, patient and hospital process variables and presence of implants. The healthcare space is very tricky, because its not just about the model and the results, its the stigmas, the hospital policies and variables and the risk factor. It will take hospitals a long time to trust these automations. Moreover there will always be the question of if the tool can match the years of experience and human expertise.

Approach
managed to achieve 71% accuracy. Sensitivity was a really important metric, especially in the field of healthcare, we want lesser false negatives, because it's always better to be on the safer side. We initially developed a classification model that proved to us that the model is indeed learning, shown through salient mapping. We then developed an object detection deep learning model using YoloV5, trained with 158 bounding-box-annotated pelvic X-ray images.