ARTIFICIAL INTELLIGENCE IN MEDICAL DIAGNOSTICS: APPLICATIONS, LIMITATIONS, AND GOVERNANCE

Abstract

Artificial intelligence (AI), and in particular machine learning methods trained on large clinical and imaging datasets, has moved over the past decade from an experimental research topic into a clinically deployed technology across multiple diagnostic specialties. Trained algorithms have demonstrated the ability to detect diabetic retinopathy from retinal photographs, identify suspicious lesions on dermatological images, flag likely malignant nodules on chest computed tomography, and detect arrhythmias from electrocardiogram signals, in several cases achieving diagnostic accuracy comparable to that of experienced specialist clinicians on standardised benchmark datasets [2–4]. These developments are driven by the convergence of three enabling factors: the increasing availability of large, digitised clinical datasets; substantial increases in computational power, particularly graphics processing units suited to training deep neural networks; and algorithmic advances in deep learning architectures specifically adapted to image and signal analysis.