Development and Validation of a Machine Learning System for Analysis and Radiological Diagnosis of Digital Chest X-ray Images

Abstract

Introduction: Medicine is identified as one of the most promising field of application for Artificial Intelligence (AI). The current research presents development and validation of a machine learning system to diagnose chest x-ray images with higher accuracy. Methodology: It was a multi-cantered, experimental study conducted from 01 July, 2021 to 30 June, 2022. The experiment was a two-step process; in the first step, a machine learning system (MLS) was developed through training, testing and tuning a specialised computer hardware & software utilising 5600 chest X ray images from NIH (National Institute of Health) chest X ray dataset. In the second step, 500 unseen chest X ray images from study centres were allowed to be diagnosed by the machine learning system and results were compared with expert opinions. Result: After the system was developed, validation was done on 3 different variations of Deep Residual Network and tested for their accuracy in classifying the findings. Using ResNet50V2, an average accuracy of 84.37%% was achieved. With case-specific variation, highest accuracy was 94.84%, highest specificity was 97.23% and highest sensitivity was 88.25%. Conclusion: With utilisation of this machine learning system, a faster radiological diagnosis of huge number of X ray images will become possible using only a small computer. Dependency on manpower, logistic support as well as rate of human-made errors can be minimised. However, this machine is never meant to replace an expert human opinion and it can never think beyond the box.