Late Breaking Abstract - Use of AI in accurate diagnosis of TB using Medical Imaging: a promising result from UAE

Abstract

Introduction: Dubai Health Authority and Agfa HealthCare recognized the potential of Machine Learning Algorithms and AI enabled workflows in medical imaging three years ago. With a strategic goal for workflow automation and fast access to diagnostic imaging results, an approach to enable Augmented Intelligence in medical imaging was devised to consider application of AI in Chest X-Ray screening.

Aim: To compare sensitivity and specificity of AI technology in screening Tuberculosis across 20 Medical Fitness Centers in Dubai, in 2015 to validate AI enabled automated Chest X-Ray screening workflow.

Method: The DHA provided Agfa HealthCare anonymized Chest X-Rays samples, half of which were categorized as Normal X-Rays, and remaining half with Tuberculosis findings based on lab confirmation. Agfa HealthCare and VRVis Vienna analyzed these anonymized X-Rays between 2015-2016 and developed a workflow concept with Machine Learning Algorithm post image processing that include edge detection and image segmentation.

Results: At the time of publication of this report (February 2018), over 4,500 Chest X-Rays have been analyzed by the AI Algorithm currently deployed at one of the DHA Medical Fitness Centers in Dubai. Dubai Validation Phase One (2017): Sensitivity: (TB Positive) 90%, Specificity: (True Negative) 55%-70%, AUC*: 0.911 (A). Phase Two (2018:re-trained algorithm) , Sensitivity: (TB Positive) 95% , Specificity: (True Negative) 70% -75%, AUC*: 0.923 (A).

Conclusion: Based on the analysis of results so far, and how the AI Algorithm is performing, cases that are flagged for a disease like Tuberculosis would get followed up on the same day