Yulei Jiang
University of Chicago, IL·Mathematics & Statistics
How do you prove that a radiologist, an imaging device, or a new diagnostic algorithm is actually good at telling disease from no disease? The standard answer is receiver operating characteristic (ROC) analysis, a statistical framework that summarizes performance across the full tradeoff between catching true cases and raising false alarms.
- Fields
Mathematics & Statistics
Medicine & Clinical Research
Biomedical Engineering
- Topics
receiver operating characteristic
ROC analysis
weighted ROC
- Grants
- 1 known grant, through Dec 2027
What they work on
- Derive weighted ROC estimation algorithms that assign a weight to each case, covering the non-parametric, conventional binormal and proper binormal models, using maximum-likelihood derivation and validation against Monte Carlo simulations.
- Build and release weighted ROC software, including an open online calculator, so that other investigators can apply case weighting to their own diagnostic performance data.
- Develop and validate a weighted ROC application that lets studies using different, non-random case samples be compared on a common footing.
- Test whether deliberately stratified case sampling followed by weighted ROC reweighting recovers the performance that random sampling would have given, enabling more efficient reader study designs.
- Estimate group diagnostic performance by averaging individual readers' ROC curves weighted by each reader's clinical case volume, and quantify inference bias from non-random case samples in multi-reader multi-case studies.
More about the research
It underpins how medical imaging technologies are evaluated. But it rests on an assumption that rarely holds: that the cases in a study are a random sample of the patients a clinician will actually see. Laboratory reader studies usually enrich for hard or interesting cases instead, and the resulting performance estimates can be biased when applied to clinical practice.
Why it matters
Decisions about which imaging devices, algorithms and screening technologies reach patients depend on reader studies whose case sets are rarely representative of real clinical populations. A weighting method that corrects for this could make laboratory results translate more faithfully to clinical performance, let investigators design smaller and better-targeted studies, and give regulators and researchers a shared, freely available tool for evaluating diagnostic technologies.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
David Chan
University of California-Berkeley, CA·Social Science
Validate data-driven measures of pneumonia diagnoses and outcomes among 5.5 million chest X-ray visits in Veterans Health Administration emergency departments. Use quasi-random assignment of patients to radiologists to infer differences in false-positive and false-negative rates.
Michael King
University of Massachusetts Chan Medical School, MA·Biomedical Engineering
Create Monte Carlo simulations of Ac-225 PSMA SPECT/CT imaging in targeted alpha therapy to jointly optimize collimator design and physics-based image reconstruction for tumor detection and quantification.
Kevin Liang
Adelphi University, NY·Physics
Derive Fisher Information matrices, a statistical measure of how precisely image features can be estimated, for imaging systems with off-axis aberrations to test whether they outperform aberration-free systems.
The school
University of Chicago: 39 labs on PI Pages, $681M a year on research, 457 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.