Computing that makes medical images answer clinical questions

The performance of a medical imaging system should be measured by how well it enables the clinical task it was built for — detecting a lesion, or quantifying a physiological parameter. Physical fidelity measures such as mean square error, signal-to-noise ratio, or structural similarity do not reliably correlate with diagnostic performance. My work develops the computational tools that make objective assessment tractable for modern, high-dimensional, nonlinear imaging systems.

With collaborators, I have established how the Bayesian ideal observer — the optimal task-based performance bound — can be approximated by Markov-chain Monte Carlo methods that use deep generative models to represent realistic object statistics; shown how learned ideal and Hotelling observers can be trained directly from simulated ensembles to estimate upper bounds on system performance, including for multimodal data; developed methods to compute and characterize the null space of an imaging operator, which determines precisely what object information a system can and cannot recover, and therefore what no reconstruction method can recover; and derived Bayesian Cramér–Rao bounds for quantitative estimation tasks, giving a principled objective for imaging system design.

Although developed in the context of photoacoustic and ultrasound imaging, these tools are modality-agnostic and apply to the assessment of any computed imaging system.

Deep learning methods routinely report large gains over classical and model-based reconstruction, but those gains are typically demonstrated with physical fidelity measures on limited test sets. My work addresses the resulting credibility gap from both directions: by exposing the specific failure modes of learned reconstruction, and by designing learned methods whose behavior can be interpreted rather than merely benchmarked.

We have shown that ill-posed tomographic problems admit multiple data-consistent solutions within the manifold of a deep generative model — a concrete mechanism for hallucination, and a tool for exploring it. We have critically assessed whether diffusion-model posterior samplers actually deliver calibrated uncertainty quantification for Bayesian inverse problems; they frequently do not, despite being widely presented as if they do. On the design side, we develop task-informed learned full-waveform inversion, in which the downstream task enters training so that reconstruction optimizes clinically relevant rather than pixel-wise performance; learned filtered-backprojection operators with analyzable structure, for which stability and consistency properties can be established; and learned forward-model and aberration corrections that remain anchored to the underlying physics.

Task-based assessment requires an ensemble of statistically realistic objects with known ground truth — something no clinical dataset can supply, because ground truth is unavailable and anatomical variability is not controllable. Virtual imaging trials close that gap, and are the practical enabler of everything above.

My long-term goal is to establish end-to-end virtual imaging frameworks as the first step in developing and validating new imaging technologies. That means constructing virtual patient populations via stochastic numerical phantoms that capture realistic variability in anatomy, tissue composition, and acoustic and optical properties; simulating the data acquisition process at high fidelity, including instrument-specific physics; applying both model-based and learned reconstruction within a common, controlled pipeline; and reading out image quality with clinically relevant, task-based measures. I have led or co-led the development of 3-D stochastic numerical breast, head, and mouse phantoms for ultrasound and photoacoustic computed tomography, and released these resources to the community. The framework is what allows learned and classical methods to be compared on identical object ensembles, at sample sizes sufficient for statistically meaningful conclusions.

Credible assessment of learned reconstruction requires a defensible reference: a model-based method that exploits the measurement physics as fully as computation allows. My work establishes such reference methods for emerging acoustic and optical modalities, and in doing so advances the quantitative accuracy achievable without learning.

This includes accurate imaging operators — a 3-D full-waveform-inversion forward model incorporating elevation-focused transducer properties for ultrasound computed tomography, and a multiphysics finite-element full-wave model for transcranial photoacoustic imaging that accounts for skull-induced aberration; novel object representations for dynamic imaging, including tensor decompositions and neural fields as a memory-efficient, self-supervised representation for spatiotemporal reconstruction; spatiotemporal methods that enable high-frame-rate dynamic photoacoustic imaging from sparse rotating-gantry measurements; and rigorous analysis of the identifiability limits of joint estimation problems.

Image reconstruction is an inverse problem, and Bayesian inverse theory provides a rigorous framework for quantifying posterior uncertainty, measuring information content, and assessing experimental-design objectives. My work develops scalable algorithms for infinite-dimensional Bayesian inverse problems and — equally important — delivers them as software the community can actually use.

I am the lead developer of hIPPYlib, an extensible framework for large-scale PDE-constrained inverse problems that has become a widely adopted reference implementation for scalable deterministic and Bayesian inversion, and a co-developer of hIPPYlib-MUQ (coupling scalable inversion with advanced MCMC) and SOUPy (PDE-constrained optimization under high-dimensional uncertainty). Methodologically, I have contributed derivative-informed reduced neural operators and projected neural networks that give fast, accurate surrogates for high-dimensional parameter-to-observable maps — an enabling technology for the many forward solves that ensemble-based assessment requires — along with Hessian-based adaptive quadrature, scalable samplers for spatially correlated random fields, and preconditioners for large-scale inversion.

These methods have been applied across imaging, computational oncology (predictive digital twins with quantified uncertainty for patient-specific decision making), geophysics, and glaciology, which has repeatedly sharpened the algorithms.

During my doctoral program and postdoctoral training I worked on large-scale numerical simulation as a tool to inform decision-making under uncertainty: patient-specific computational hemodynamics to quantify wall shear stress and predict aneurysm rupture risk; optimal control of the inlet condition of a turbulent jet to ensure proper mixing; information-theoretic approaches to optimally designing sensing systems; and algebraic multigrid and numerical upscaling for flow in porous media.