Machine learning for classifying hip fracture
Given the high mortality rates and costs for hip fractures, any improvement in classification and hence treatment will have significant benefits. Led by Prof. Richie Gill, the aim of this research was to work with the IMI MIRAs to develop a machine learning (ML) based method to automatically classify hip fractures using X-rays, with the eventual aim of standardising classification. This was done in 2 stages, firstly to automatically detect the hip joint(s) in a radiograph, and then classify the fracture given this selected region.
A fully convolutional network was trained to automatically locate the hip joint, using manually annotated X-ray images. Transfer learning of a pre-trained convolutional neural net was then used to automatically classify the type of fracture using a data set of almost 22 thousand images. The algorithms must be robust enough to cope with variability of image quality and patient positioning and orientation in the X-Ray image. In parallel, each fractured hip was classified (no fracture, intertrochanteric and intracapsular) by at least 2 experts, with 2 further reviewers deciding the class when there was disagreement.
Finding the hip was 99.2% accurate. Human experts agreed on only 59% of cases. The overall accuracy of fracture classification by the ML algorithm was 92%, the accuracy varied between classes, 94% for no fracture, 91% for intertrochanteric and 89% for intracapsular. The machine learning method was more accurate in classifying fracture than human observers.
Together, the research team has established a proof of concept that more accurate and standardised methods for hip fracture classification can be developed integrating a combination of Machine Learning technology in collaboration with consultant expertise.