Artificial intelligence applications for the acquisition, analysis and reporting of cardiovascular magnetic resonance imaging

Abstract

From the moment a patient enters the scanner to have a cardiovascular magnetic resonance (CMR) scan, a clinical workflow is set in motion. After image acquisition, quality assurance and analysis, the ultimate product of this workflow is a text report of the scan findings that will (hopefully) answer clinical questions unambiguously, and help progress the patient’s care. Typically, the following steps occur: • A ‘protocol’ of imaging sequences is acquired based on the clinical question. • Quality assurance is performed by Radiographers ‘eyeballing’ images as they are produced by the scanner. • Initial images are reviewed to plan later ones and determine whether initial findings require bespoke imaging. • The patient leaves the scanner and images are transmitted to a server to be retrieved by the reporting clinician. • The clinician performs analyses including visual inspection and manual measurement of dimensions, volumes and blood flow. • The clinician synthesises their analyses into a free-text report of scan findings, which is uploaded to the electronic health record. • The images and report are used by the referrer to determine the patient’s treatment. They are also stored for future use in audit, quality improvement and research. Every stage of this workflow draws on humans’ skills, knowledge and experience; but there is also the potential for inaccuracy, inconsistency, system inefficiency, uncertainty and clinical equipoise. At each stage there is an interaction between human and computer; each interaction presents an opportunity for automation. Artificial intelligence (AI) is a type of programming that enables computers to perform tasks automatically, without being explicitly programmed to do so. The core aim of this thesis is to develop and apply the latest AI technology throughout the clinical CMR workflow. In this thesis I explore different AI applications to provide automated assistance for: 6 • Acquisition - Rapid interpretation of early anatomy images to aid radiographer decision-making (Chapter 4). • Quality assurance - Real-time assessment of late gadolinium enhancement likelihood to prompt further imaging for uncertainty-reduction (Chapter 5). • Analysis & Reporting – Automatic removal of slices that do not image the left ventricle to reduce clinician distraction when reporting (Chapter 6) • Research & Audit – Automatic classification of clinical diagnoses from text reports (Chapter 7). Each of these is an example of a task in the clinical CMR workflow that is either done inefficiently or inaccurately by humans, or not done by humans at all because doing it manually is too difficult or time-consuming; this makes it amenable to automation using AI. For each task I identified a type of AI, adapted it, and applied it to the specific problem. All the applications have different structures and architectures that I have selected and adapted to handle each task. Finally, in a unified synthesis, I bring together the common themes arising from these applications and discuss them within the wider context of clinical AI development, challenges for adoption, cultural barriers, ethical considerations, and the future of clinical training.