Artificial Intelligence in Cardiac Magnetic Resonance Imaging to Predict Prognosis and Treatment Response

Background Pulmonary arterial hypertension (PAH) is a serious disease of the heart and lungs. Its impact on patients can be severe, including limitation of day-to-day activities and high mortality. The diagnosis, treatment and monitoring of PAH are challenging and there is a need for tools that can aid clinical decision-making to optimise patient outcomes. Cardiac MRI (CMR) provides both qualitative and quantitative information about cardiac function and is an important method for evaluating the severity of PAH. The application of machine learning (ML) tools is of growing interest in medical imaging. ML has the potential to automate complex and repetitive tasks, including the rapid segmentation of anatomical structures on images and extraction of clinically useful information. Aims This thesis proposes the combination of CMR with two different ML tools to predict prognosis and treatment response in PAH. The first ML tool involves the automated measurement of different cardiac parameters and assesses their utility in predicting prognosis and treatment response. The second ML tool involves the extraction of imaging features directly without the need for segmentation to predict the risk of mortality. My Contribution The ML models in this thesis were developed at the University of Sheffield in collaboration with Leiden University. Sheffield is a centre of excellence in PAH treatment thanks to the Sheffield Pulmonary Vascular Disease Unit, which is one of the largest internationally. Each year, more than 700 PAH patients undergo CMR for diagnosis and monitoring. Additionally, each newly diagnosed patient has accompanying in-depth clinical phenotypic data, including right heart catheterisation, exercise and pulmonary function tests, and quality of life assessment. During my research, I created and curated a dataset combining imaging and time-matched clinical data. I identified eligible CMR scans, landmarked and contoured cardiac chambers on multiple sequences and organised the collaboration with computer scientists at Leiden and Sheffield. I arranged image anonymisation, storage and transfer and advised computer scientists on the clinical relevance of CMR images. I performed quality control on ML analyses, collated their results, and analysed the data within clinical context. I have written all chapters in this thesis and clarified the roles of my co-authors at the end of each chapter. Thesis Outline Chapter 1 provided an overview of the growing role of CMR in the diagnosis and evaluation of PAH. Chapter 2 summarised the prognostic value of CMR measurements in the prediction of clinical worsening and mortality in PAH patients. Chapter 3 illustrated the rapid expansion of research using AI approaches to automate CMR measurements. The quality of the existing literature was reviewed, significant shortcomings in the transparency of studies were identified and solutions were recommended. Chapter 4 showed our experience in developing, validating and testing a fully automatic CMR segmentation tool. Our tool was developed in one of the largest multi-vendor, multi-centre and multi-pathology reported datasets, and included a large group of patients with right heart disease. We implemented the lessons learned in Chapter 3 and provided extensive descriptions of our datasets, ML model and performance. Our model showed excellent reliability, generalisability, agreement with CMR experts and correlation with invasive haemodynamics. Chapter 5 demonstrated that the automatic CMR measurements allowed assessment of patient-orientated outcomes and prediction of mortality. Thresholds of changes in CMR metrics were identified that could inform clinical decisions in the monitoring of PAH patients. Chapter 6 showed promising results of an ML tool to extrapolate prognostic CMR features with incremental value compared to clinical risk scores and volumetric CMR measurements. Finally, Chapter 7 showed that myocardial T1 mapping could potentially add diagnostic and prognostic value in PAH. Impact and Future Direction In addition to the known advantages of ML for providing rapid results with minimal human involvement, the ML tools developed in this thesis allow visualisation of outcomes and are transparent to the human assessor. ML applications to automate the measurement of CMR metrics and extract prognostic imaging features have potential to add clinical value by (i) streamlining prognostication, (ii) informing treatment selection, (iii) assisting the monitoring of treatment response and (iv) ultimately improving clinical decision-making and patient outcomes. Additionally, these tools could point to new CMR end-points for clinical trials, accelerating the development of new treatments for PAH. ML will likely elevate the role of CMR as a powerful prognostic modality in the years to come. Looking ahead, I hope to combine multi-source clinical, imaging and patient-orientated data from several ML tools into a single package to facilitate the assessment of cardiovascular disease.