BioPhysics and AI

HeartFlow · Stanford · UCSD · Cornell · 2004–2019

For fifteen years I combined physics-based simulation with machine learning to understand biological systems, from blood flow in the human heart to memory circuits in the brain. Several of these methods are now used routinely in the clinic.

Blood flow · Cardiac imaging · Neuroscience · Uncertainty quantification

HeartFlow · UCSD

Real-time, physics-driven blood flow simulation

Blood flow is a key indicator for the prognosis and treatment of cardiovascular disease, and hemodynamic measurements have become the gold standard for choosing between medical therapy and stenting. I developed algorithms that simulate blood flow in patient-specific models built from CT scans, using computational fluid dynamics with boundary conditions from a model of microcirculatory resistance. Solving the millions of equations that come from the Navier–Stokes equations requires high-performance computing, so we also built machine-learning and reduced-order models that match full 3D simulations in real time.

Blood flow in the aorta and coronary arteries — Quantification of blood flow in the aorta and coronary arteries.

FDA-cleared surgical stent planner

I led the R&D effort to design, develop and launch a real-time stent planner at HeartFlow. It was cleared by the FDA and is used by physicians in the United States and abroad.

Watch the planner

HeartFlow Planner

Hemodynamics estimated with machine learning — Hemodynamic parameters estimated with machine learning (left) and the time complexity of ML compared with full 3D solutions (right).

HeartFlow

AI for cardiovascular imaging

CT, MRI and ultrasound scans contain rich information for managing patient risk and treatment. HeartFlow built one of the largest databases of human coronary artery trees with clinical outcomes, and with the imaging sciences team I developed AI methods that predict anatomy and hemodynamics in real time.

Stenosis quantification. HALE (healthy area of lumen estimation) quantifies vessel narrowing across a diverse set of lesion morphologies and agrees better with expert readers than prior state-of-the-art methods. It was published at MICCAI and is used routinely to diagnose coronary artery disease.

Machine learning stenosis detection compared with CT and angiogram — ML-based stenosis detection compared against CT and angiography.

Factors that influence CT image quality — Factors that influence image quality (top) and CCTA scans with excellent and poor quality (bottom).

Automated image quality. Our image-quality model combines local intensity features, a cross-correlation mis-registration metric and global factors such as contrast-to-noise ratio. Trained on hundreds of annotated scans, it performs as well as expert radiologists (European Radiology, 2018).

Stanford University · Deisseroth Lab

Computational neuroscience: hub neurons

How are memories stored and retrieved? With Karl Deisseroth, Priya Rajasethupathy and colleagues at Stanford, I built tools to quantify neuronal activity from in-vivo two-photon imaging of behaving mice and to analyze it as a graph.

We found that memory is encoded as a small-world network, with highly connected hub neurons that are preferentially recruited during retrieval. The work was published in Nature with me as co-first author.

Fluorescence extracted from two-photon images — Fluorescence extracted from in-vivo two-photon images of behaving mice.

Highly connected hub neurons in the hippocampus — Highly connected hub neurons revealed by imaging the hippocampus of mice in different contexts.

Cornell University

Uncertainty quantification and stochastic modeling

My Ph.D. thesis developed methods for solving partial differential equations when geometry, boundary conditions or parameters are uncertain, for systems such as flow in porous media, heat conduction and manufacturing processes. The unknowns vary over space, time and a stochastic space. Stochastic models are built with the maximum entropy principle and optimized with stochastic gradient descent.

The same ideas later quantified how uncertainty in segmentation, boundary conditions and blood viscosity affects cardiovascular simulations.

Framework for stochastic PDEs in porous media — Solving and optimizing stochastic PDEs for flow through porous media.

Information learning framework for microstructure models — An information learning framework (stochastic machine learning with a Kullback–Leibler divergence error measure) for generating stochastic microstructure models.

Selected publications

Rajasethupathy, P.*, Sankaran, S.*, Marshel, J.H.*, et al. Projections from neocortex mediate top-down control of memory retrieval. Nature 526, 653–659, 2015. (*equal contribution) [paper]
Sankaran, S., Lesage, D., Tombropoulos, R., Xiao, N., Kim, H.J., Spain, D., Schaap, M. and Taylor, C.A. Physics driven real-time blood flow simulations. Computer Methods in Applied Mechanics and Engineering 364, 112963, 2020. [paper]
Modi, B.N., Sankaran, S., et al. Predicting the physiological effect of revascularization in serially diseased coronary arteries. Circulation: Cardiovascular Interventions 12(2), 2019. [PDF]
Nakanishi, R., Sankaran, S.*, et al. Automated estimation of image quality for coronary CT angiography using machine learning. European Radiology 28(9), 4018–4026, 2018. [PDF]
Sankaran, S., et al. HALE: Healthy area of lumen estimation for vessel stenosis quantification. MICCAI, LNCS 9902, 380–387, 2016. [PDF]
Sankaran, S., Grady, L. and Taylor, C. Fast computation of hemodynamic sensitivity to lumen segmentation uncertainty. IEEE Transactions on Medical Imaging 34(12), 2562–2571, 2015. [PDF]
Sankaran, S., Moghadam, M.E., Kahn, A.M., Tseng, E.E., Guccione, J.M. and Marsden, A.L. Patient-specific multiscale modeling of blood flow for coronary artery bypass graft surgery. Annals of Biomedical Engineering 40(10), 2228–2242, 2012. [paper]

All publications →