Lianrui Zuo

Research

Medical images are measurements, and measurements come with an instrument attached. My work is organized around three questions about what a measurement can and cannot tell us. The papers below are a selection; the full list is under Publications.

See the patient, not the scanner

The same brain scanned on two machines gives two different images. Before anything downstream can be trusted (a volume, a segmentation, a risk score), the variation that comes from the instrument has to be separated from the variation that comes from the patient. Much of my training went into this problem, mostly through disentangled representation learning for MRI, and it now extends to CT reconstruction kernels and to studies that span more than a hundred scanners.

Selected papers

Recover what was not captured

Clinical images are rarely complete. A chest CT is cropped to the lungs, a contrast is missing from a protocol, a study keeps a handful of 2D slices where a 3D volume was needed. I use generative models, recently latent diffusion and flow matching, to recover what the scanner did not capture: extending a truncated field of view, building 3D volumes from a few 2D slices, filling and synthesizing lesions, and synthesizing contrasts that were never acquired. The interesting part is not only making the missing region look right, but knowing what the recovered region can and cannot be used for.

Selected papers

Ask what each measurement adds

A patient accumulates scans, time points, and records, and each one costs something to acquire, process, and interpret. I want to quantify what each of them actually adds to a clinical prediction, rather than assume that more data is always better. Lung cancer screening is where I am starting; it is the focus of my NIH/NCI K99/R00 award.

Selected papers

Funding