Lianrui Zuo

Research

Medical images are measurements, and every measurement comes from a device or scanner. My work is organized around three questions about what a measurement (e.g., imaging data or electronic health records) can and cannot tell us.

Where each question sits
Patient 1 2 3
  1. 1See the patient, not the scanner the device or scanner leaves its mark on the measurement
  2. 2Recover what was not captured a measurement holds more than what it shows
  3. 3Ask what each measurement adds does the measurement change the decision?
1

See the patient, not the scanner

Two patients, five scanners. Each scanner draws the same anatomy with its own contrast, resolution, and noise. Harmonize, and the columns stop differing while the two rows still do.

Measurements of the same patient taken on different devices or scanners can differ, and this undesired variation often biases the downstream interpretation of medical data. I develop harmonization methods to remove the undesired scanner variation in MRI and CT while retaining the true biological signal from the patient.

Selected papers
2

Recover what was not captured

Inside the frame is what was measured. Recover infers the patient just around it, and the lines fade as the inference gets less certain. It stops there: one measurement cannot tell you the whole patient. Show the patient to compare, and move the frame with your cursor (on a phone, drag it).

The information in a biomedical measurement is not limited to what we can see in it directly. A chest CT, for example, is reconstructed within a limited field of view that often cuts off part of the body, yet the image still carries information about what lies just outside it. I use generative models to extend a truncated field of view, and more broadly to recover what the device or scanner did not capture, such as missing slices and contrasts. The recovered part is most reliable close to what was measured, and part of the work is knowing how far it can be trusted.

Selected papers
3

Ask what each measurement adds

Each frame is a measurement this patient could have; click one to take it. The bar shows how much of the patient is pinned down, split by what each measurement added when it came in. Frames 1 and 2 overlap, so whichever comes second adds little.

A patient accumulates measurements: images from different devices or scanners, repeated over time, alongside clinical records. Each one costs something to acquire and interpret. I want to measure what each of them adds to a clinical decision, given what is already known, rather than assume that more data is always better.

Selected papers

Funding