Please login to view this media
- Talk
- 20/09/2022
- UK
Developing an Artificial Intelligence Diagnostic Tool for Paediatric Distal Radius Fractures, a Proof of Concept Study
Description
In this presentation, Sriharsha Aryasomayajula, a machine learning engineer at Woodside, discusses an innovative project aimed at developing an artificial intelligence diagnostic tool for pediatric distal radial fractures. The project, a collaboration between St George University Hospital and Kingston University, addresses a significant issue in the UK where one in 50 children sustain a fractured bone annually, with many injuries not visible on initial X-rays. The economic burden of these injuries is staggering, costing the healthcare system around £2 billion each year and incurring substantial litigation costs.
Sriharsha outlines the current landscape of pediatric distal radial fracture studies, highlighting the potential of using ultrasound images combined with AI to enhance diagnosis. However, he notes the limitations posed by the current lack of quality datasets for training predictive algorithms. The objective of the project is to create a machine learning model capable of accurately identifying fractures in radiographs.
The methodology involves labeling radiographs, training a convolutional neural network, and ensuring ethical standards through the TRIPOD statement checklists for diagnostic models. The project employs a dataset of 5,000 pediatric wrist radiographs collected over several years, focusing on high-quality images to maximize model accuracy.
Sriharsha shares insights on the data preprocessing, which includes de-identifying patient information and converting image files into a compatible format for neural network processing. The team applies image augmentation techniques to expand the dataset, ultimately achieving a training accuracy of 96% and a diagnostic test accuracy of 85%.
He concludes by emphasizing the necessity for robustness in the AI model to cater to various noise types and expresses a desire to develop an explainable AI that can visually indicate fracture locations within radiographs.