| Summary: | Pulmonary hypertension (PH) is a clinical condition characterised by an increased mean pulmonary arterial pressure (mPAP) of over 25 mmHg measured, at rest, by right heart catheterisation (RHC). RHC is currently considered the gold standard for diagnosis, follow-up and measurement of response to treatment. Although the severe complications and mortality risk associated with the invasive procedure are reduced when it is performed in a specialist centre, finding non-invasive PH diagnosis methods is highly desirable. Non-invasive, non-ionising imaging techniques, based on magnetic resonance imaging (MRI) and on echocardiography, have been integrated into the clinical routine as means for PH assessment. Although the imaging techniques can provide valuable information supporting the PH diagnosis, accurately identifying patients with PH based upon images alone remains challenging. Computationally based models can bring additional insights into the haemodynamic changes occurring under the manifestation of PH. The primary hypothesis of this thesis is that that the physiological status of the pulmonary circulation can be inferred using solely non-invasive flow and anatomy measurements of the pulmonary arteries, measured by MRI and interpreted by 0D and 1D mathematical models. The aim was to implement a series of simple mathematical models, taking the inputs from MRI measurements, and to evaluate their potential to support the non-invasive diagnosis and monitoring of PH. The principal objective was to develop a tool that can readily be translated into the clinic, requiring minimum operator input and time and returning meaningful and accurate results. Two mathematical models, a 3 element Windkessel model and a 1D model of an axisymmetric straight elastic tube for wave reflections were implemented and clinically tested on a cohort of healthy volunteers and of patients who were clinically investigated for PH. The latter group contained some who were normotensive, and those with PH were stratified according to severity. A 2D semi-automatic image segmentation workflow was developed to provide patient specific, simultaneous flow and anatomy measurements of the main pulmonary artery (MPA) as input to the mathematical models. Several diagnostic indices are proposed, and of these distal resistance (Rd), total vascular compliance (C) and the ratio of reflected to total wave power (Wb/Wtot) showed statistically significant differences between the analysed groups, with good accuracy in PH classification. A machine learning classifier using the derived computational metrics and several other PH metrics computed from MRI images of the MPA and of the right ventricle alone, proposed in the literature as PH surrogate markers, was trained and validated with leave-one-out cross-validation to improve the accuracy of non-invasive PH diagnosis. The results accurately classified 92% of the patients, and furthermore the misclassified 8% were patients with mPAP close to the 25 mmHg (at RHC) threshold (within the range of clinical uncertainty). The individual analysis of all PH surrogate markers emphasised that wave reflection quantification, although with lower diagnosis accuracy (75%) than the machine learning model embedding multiple markers, has the potential to distinguish between multiple PH categories. A finite element method (FEM) based model to solve a 1D pulmonary arterial tree linear system, has been implemented to contribute further to the accurate, non-invasive assessment of pulmonary hypertension. The diagnostic protocols, including the analysis work flow, developed and reported in this PhD thesis can be integrated into the clinical process, with the potential to reduce the need for RHC by maximising the use of available MRI data.
|
|---|
