1.Background:
Flow biomarkers, such as pressures inside the heart and great vessels, are central to diagnosing and managing cardiovascular disease and heart failure; however, they are mostly measured by invasive catheterisation. Image-based computer simulations of blood flow could deliver these biomarkers non-invasively; yet current estimates of flow rely on tailored computational simulations that can take days to set-up and run. AI has shown potential for simulating flow rapidly in toy problems but is unable to translate this success to real-world data, which is dynamic, non-linear, noisy and highly specific to the anatomy of each patient [1][2].
2.Novelty and Importance:
In this PhD, we address these challenges through developing a novel framework for physics-constrained AI for blood flow prediction that uses Volumetric Integrated Spatio-Temporal Attention (VISTA-Flow). VISTA-Flow can learn to simulate complex flow from Cine MRI data in ways that robustly generalise across individuals. These capabilities will open the door to tailored simulations (Digital Twins) of the hearts of individual patients, from which earlier diagnosis and in-silico testing of treatments will become possible.
3.Aim and Objectives:
Our goal is to develop, train and test VISTA-Flow in a range of problems with increasing complexity, to demonstrate feasibility on 3D patient-specific cardiac anatomies. This work will build on our existing personalised finite-element models of cardiac flow mechanics (De Vecchi, [3]) and multi-scale vision transformer MS-SiT (Robinson, [4]) and will be structured into three aims: (i) enhance and validate MS-SiT on 2D analytical flow problems; (ii) extend it to anatomically realistic 2D settings with moving walls and valves; (iii) scale to 3D patient-specific anatomies of heart failure patient with validation against 4D Flow MRI and CFD simulations. The resulting VISTA-Flow will finally be validated on prospective clinical data for proof of concept.
[1] Luo H, Wu H, Zhou H, Xing L, Di Y, Wang J, Long M. Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries. 2025;
[2] Lu L, Jin P, Karniadakis GE. DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. Nat Mach Intell Nature Research; 2020;3:218–229.
[3] de Vecchi A, Gomez A, Pushparajah K, Schaeffter T, Simpson JM, Razavi R, Penney GP, Smith NP, Nordsletten DA. A novel methodology for personalized simulations of ventricular hemodynamics from noninvasive imaging data. Comput Med Imaging Graph 2016;51:20–31.
[4] Dahan S, Fawaz A, Williams LZJ, Yang C, Coalson TS, Glasser MF, David Edwards A, Rueckert D, Robinson EC, Fawaz D, Coalson WY, Edwards G, Robinson R. Surface Vision Transformers: Attention-Based Modelling applied to Cortical Analysis. Proc Mach Learn Res 2022;172:1–22.