Characterization of left atrial hemodynamics with 4D flow magnetic resonance imaging, in-silico models and deep learning

Abstract

Although the left ventricle has traditionally held the spotlight as the primary determinant of cardiovascular disease prognosis, the characterization of left atrial hemodynamics is increasingly being recognized as a potential bellwether of several cardiovascular disorders. However, blood flow within the left atrium is pulsatile, multidirectional, and time-varying in nature, and imaging modalities currently available in clinical practice can only provide limited insight into the intricate underlying hemodynamic patterns. As a result, this thesis aims to address the challenges in the implementation of cutting-edge cardiac hemodynamic analysis methods, including 4D Flow magnetic resonance imaging (4D Flow MRI), computational fluid dynamics, and deep learning, to standardize and expedite the characterization of left atrial hemodynamics, laying the foundation for a more comprehensive understanding of the role of left atrial blood flow in cardiovascular disease. To achieve this objective, the thesis is structured around three primary contributions. Firstly, a robust computational framework was developed for the analysis of 4D Flow MRI in the left atrium, capable of handling multicenter data with varying quality. The framework is completely open-source and facilitates a comprehensive qualitative and quantitative analysis of advanced hemodynamic parameters that were previously unexplored in any sizeable left atrial cohort. Secondly, we derived pulsed-wave Doppler-like velocity spectrograms from 4D Flow MRI to study its potential use in left ventricular diastology as an adjunct in challenging cases where standard echocardiographic assessment may not be feasible. Finally, to address the need for real-time monitoring of patient-specific hemodynamics in the clinical setting, we developed a set of deep learning surrogates, leveraging recent advances in graph and physics-informed neural networks to enable adaptation to unstructured data and surpass the limitations of the current data-driven paradigm.

Tradizionalki, ezker bentrikulua gaixotasun kardiobaskularren pronostikoaren lehen determinatzaile foko nagusia izan bada ere, gero eta gehiago antzematen da ezker atrioko hemodinamikaren karakterizazioaren potentziala hainbat gaixotasun kardiobaskularren adierazle bezala. Hala ere, ezkerreko atrioko odol-fluxua pulsatila, norabide anitzekoa eta denbora-aldakorra da, gaur egun praktika klinikoan eskuragai dauden irudi-modalitateek ganbera barruan aurkitzen diren patroi hemodinamiko konplexuen irudi oso mugatua eskaini dezaketen bitartean. Ondorioz, tesi honen helburu nagusia bihotzeko analisi hemodinamirako punta-puntako metodoen (4D fluxu erresonantzia magnetikoa (4D Flow MRI), fluidoen dinamika konputazionala eta ikaskuntza sakona barne) implementazioan dauden erronkei aurre egitea da, ezker atrioko hemodinamikaren karakterizazioa estandarizatu eta bizkortzeko eta gaixotasun kardiobaskularren garapenean duen garrantzia hobeto ezartzeko. Helburu hori lortzearren, tesia hiru ekarpen nagusiren inguruan egituratzen da. Lehenik eta behin, 4D Flow MRI ezker atrioan aztertzeko kodigo irekiko framework konputazional bat garatu da kalitate ezberdinareko datu multizentrikoak maneiatzeko gai dena. Framework honek, atrio azkerrean aurretik aztertu ez diren parametro hemodinamiko eratorrien analisi kualitatibo eta kuantitatiboa ahalbidetzen du. Bigarrenik, Doppler pultsatuaren tankerako abiadura-espektrogramak garatu ditugu 4D Flow MRI, ebaluazio ekokardiografiko estandarra bideragarria ez den kasuetan ezkerreko diastologia bentrikularrean izan dezakeen erabilera aztertzeko. Azkenik, ingurune klinikoan hemodinamika denbora errealean monitorizatzeko beharrari aurre egiteko, ikaskuntza sakoneko modelo multzo bat garatu dira, fisikan-informatutako eta grafoen sare neuronaletan eman diren aurrerapenak baliatuz, datuetan funtsatutako egungo paradigmaren mugak gainditzeko eta datu ez-estrukturatuetara egokitzeko.