Prince Lab
Prince LabQuantitative MRI & imaging AI

Department of Radiology · Weill Cornell Medicine

Routine MRI, measured well enough to follow disease.

We build and validate deep-learning tools that turn the abdominal MRI patients already receive into organ, cyst and body-composition measurements. The goal is a number that holds steady from one scan to the next, so a change on follow-up means the disease changed. Much of that work centers on autosomal dominant polycystic kidney disease (ADPKD).

“Prince Lab” is a working name. Swap in the lab's preferred title.
Explore the research cores Publications
Kidney volume test–retest difference when a 3D model averages five MRI sequences, better than all five expert readers
1.3%
He et al., Acad Radiol 2024
Less expert contouring time with the deployed kidney segmentation model (1,724 s to 723 s per case)
51%
Goel et al., Radiol AI 2022
Dice similarity of kidney segmentation on external validation in 20 patients from outside institutions
0.98
Goel et al., Radiol AI 2022
Individual cysts contoured on every available scan, mean follow-up 11 years (1,654 contours)
299
Hu et al., Commun Med 2026

Principal investigator

Martin R. Prince, MD, PhD

Professor of Radiology, Weill Cornell Medicine Adjunct Professor of Radiology, Columbia University Vagelos College of Physicians and Surgeons Attending Radiologist, Body Imaging, NewYork-Presbyterian

Dr. Prince trained at Harvard Medical School and MIT, interned in medicine at UCSF, and completed his radiology residency and fellowships in angiography and MRI at Massachusetts General Hospital. At MGH in 1993 and 1994, he showed that a gadolinium infusion timed to a 3D acquisition could make arteries bright while veins and background tissue enhanced only minimally. Contrast-enhanced MR angiography grew out of that observation. At the University of Michigan he went on to develop a bolus-triggered version that GE commercialized as SmartPrep.

His group later quantified the risk of nephrogenic systemic fibrosis after gadolinium in a two-center series of more than 83,000 patients, and traced the disease through a review of 639 biopsy-confirmed cases. The lab's current work points deep learning at routine abdominal MRI, in close partnership with nephrologists at The Rogosin Institute, to make organ and cyst volumes reproducible enough to track ADPKD year to year.

Training

  • MD, Harvard Medical School
  • PhD, Massachusetts Institute of Technology
  • Internship in medicine, UCSF
  • Radiology residency; fellowships in angiography and MRI, Massachusetts General Hospital

Recognition & service

  • ISMRM Gold Medal
  • Fellow, American College of Radiology
  • Endowed the Prince Research Resident Grant, RSNA Research & Education Foundation
  • Has served as associate editor of Radiology
Add award years; confirm current editorial roles.

Books

  • 3D Contrast MR Angiography, with T. M. Grist and J. F. Debatin. Springer, 3rd ed.
  • MRI from Picture to Proton, with D. W. McRobbie, E. A. Moore and M. J. Graves. Cambridge University Press, 3rd ed., 2017

Milestones

  1. Dynamic gadolinium-enhanced 3D abdominal MR arteriography: preferential arterial enhancement in all 16 patients.

    Journal of Magnetic Resonance Imaging

  2. Gadolinium-enhanced MR aortography in 125 patients. In the 48 with angiographic or surgical correlation, 88% sensitivity and 97% specificity for stenosis or occlusion.

    Radiology

  3. NSF incidence at two centers: none among 74,124 standard-dose patients, 15 of 8,997 given a high dose.

    Radiology

  4. Systematic review of 639 biopsy-confirmed NSF cases. Only seven followed contrast given after 2008.

    Radiology

  5. Kidney segmentation model deployed into clinical use for ADPKD, halving expert contouring time.

    Radiology: Artificial Intelligence

  6. Averaging model-assisted volumes across pulse sequences more than doubles kidney volume reproducibility.

    Journal of Magnetic Resonance Imaging

  7. Kidney, liver and cyst volumes measured automatically in one pipeline.

    Journal of the American Society of Nephrology

  8. Natural history of 299 individual cysts over a mean 11 years of serial MRI.

    Communications Medicine

Research cores

Six cores, one measurement pipeline

Each core is a line of work with its own questions, but they share data, models and readers. A biomarker study in Core 1 depends on the segmentation models in Core 2 and the checks in Core 3, and reading the same exams carefully is how the findings in Core 4 turned up.

Core names and groupings are a proposal drawn from the lab's publication topics.
Core 1

ADPKD imaging biomarkers

Kidney · liver · individual cysts

Total kidney volume is how clinicians stage ADPKD, yet it sums thousands of cysts that behave differently. This core follows kidneys, livers and individual cysts across years of serial MRI, looking for measurements that flag fast progression earlier and show whether a treatment is working.

Recent work

  • Simple T2-bright cysts grow logistically, median 11% a year. Over follow-up, 42% (94 of 222) changed course: 29% shrank, 16% disappeared and about 6% turned complex or T1-bright. Commun Med 2026
  • Fitting serial volumes with two-parameter least squares improves predictions of kidney growth rate. Sci Rep 2024
  • Feasibility studies of baseline kidney growth rate as a guide to tolvaptan benefit, and of water therapy for slowing progression. J Clin Med 2025 · Kidney360 2024
  • Kidney and liver cyst growth during pregnancy, and quantitative susceptibility mapping for stones, hemorrhage and cyst type. J Clin Med 2025 · Abdom Radiol 2024

Lab authors

Zhongxiu (Sue) Hu, Hreedi Dev, Vahid Bazojoo, Arman Sharbatdaran, Xinzi He

All ADPKD biomarker papers
Core 2

Deep-learning segmentation

Models in clinical use

Contouring polycystic kidneys by hand took close to half an hour per case. The first model, a U-Net with an EfficientNet encoder developed on 213 exams from 129 patients, went into clinical use and cut that time by 51%. The pipeline has since grown to cover liver and spleen, five pulse sequences plus CT, and cysts in the liver, kidney and pancreas.

What we have built

  • Deployed kidney model: Dice 0.98 on external and 0.97 on prospective validation. Radiol AI 2022
  • Kidney, liver and spleen together, with 42% less reader time and lower measurement spread. Tomography 2022
  • 3D multimodality nnU-Net trained and validated on T1, T2, SSFP, DWI and CT from 413 subjects; Dice above 97% for all three organs. Acad Radiol 2024
  • Liver cyst segmentation, pancreatic cyst detection, and kidney, liver and cyst volumes in a single pass. Radiol Adv 2024 · Tomography 2024 · JASN 2026

Lab authors

Xinzi He, Arman Sharbatdaran, Qing (Kristina) Xiong, Hreedi Dev, Zhongxiu (Sue) Hu

All deep-learning papers
Core 3

Reproducibility & quality control

Test–retest · outlier analysis

Kidney volume from a single sequence varies by roughly ±5%, about the same as a year of ADPKD growth. Measuring every sequence the exam already contains, discarding outliers and averaging the rest brings the error down far enough to see real change between visits.

What we have found

  • Averaging five model-assisted segmentations cut scan–rescan difference to 2.5%, and to 2.1% after excluding one outlier. JMRI 2023
  • With a 3D model averaging five sequences, kidney volume test–retest difference fell to 1.3%. Acad Radiol 2024
  • Acquisition errors, mostly breathing-related slice misregistration, explained 88% of outlier measurements. Tomography 2023
  • Only 21% of 145 outside radiology reports included kidney volume; one included liver volume. Kidney Int Rep 2025

Lab authors

Hreedi Dev, Arman Sharbatdaran, Xinzi He, Zhongxiu (Sue) Hu

All reproducibility papers
Core 4

Extrarenal findings on routine MRI

Chest · heart · spine · bile ducts · IVC

An abdominal MRI ordered for kidney volume also images the lung bases, heart, spine, bile ducts and inferior vena cava. Reading those structures systematically, against matched controls, has surfaced findings that are easy to miss one scan at a time.

What we have found

  • Pleural effusion in 21% of 268 ADPKD subjects versus 8% of matched controls. J Clin Med 2023
  • Pericardial effusion thicker than 5 mm in 21% of 117 ADPKD subjects versus 3% of matched controls. J Clin Med 2022
  • Severe IVC compression by cysts in 15% of ADPKD subjects and in no controls. Kidney Int Rep 2021
  • Pancreatic cysts in 36% of 110 ADPKD patients versus 23% of matched controls, and in 62% of those with PKD2 mutations. Radiology 2016
  • Spinal meningeal diverticula, bile duct dilatation, and left ventricular hypertrophy read from abdominal MRI. AJNR 2025 · Abdom Radiol 2026 · JCAT 2026

Lab authors

Qing (Kristina) Xiong, Vahid Bazojoo, Zhongxiu (Sue) Hu, Arman Sharbatdaran, Hreedi Dev

All extrarenal papers
Core 5

Organ volumetry beyond ADPKD

Spleen · liver · muscle

Model-assisted contouring carries over to other settings where an organ or muscle measurement matters clinically: splenomegaly in myelofibrosis, liver volume before transplantation, and trunk muscle quality on CT.

Recent work

  • Fast, reproducible spleen volumes for tracking myelofibrosis progression and treatment response. J Clin Med 2025
  • MRI and CT liver volumes checked against surgical specimen weight in patients undergoing liver transplantation. JCAT 2026
  • An automated muscle pipeline linking thoracic kyphosis to trunk muscle-fat fraction on CT. Tomography 2026

Lab authors

Arman Sharbatdaran, Vahid Bazojoo, Xinzi He, Hreedi Dev

All papers in this core
Core 6

Contrast MRA & gadolinium safety

Where the lab started

Gadolinium-enhanced 3D MR angiography began with Dr. Prince's work at Massachusetts General Hospital in the early 1990s. Later studies at Weill Cornell and Columbia measured the risk of nephrogenic systemic fibrosis by dose and kidney function and tracked how that risk changed after 2008.

Key results

  • Timing a gadolinium infusion to the 3D acquisition gave strong arterial enhancement without excessive venous or background enhancement. JMRI 1993 · Radiology 1994
  • Automatically detecting contrast arrival in the aorta raised arterial enhancement from 19-fold to 28-fold over manual timing and cut venous enhancement. Radiology 1997
  • Gadolinium-enhanced pulmonary MRA in 30 patients: sensitivity 75–100% and specificity 95–100% across three blinded readers, compared with conventional angiography. N Engl J Med 1997
  • No NSF in 74,124 standard-dose patients; 15 cases (0.17%) among 8,997 given a high dose. Radiology 2008
  • Of 639 biopsy-confirmed cases in the literature, seven followed contrast given after 2008. Radiology 2019

Lead

Martin R. Prince

All MRA & gadolinium papers

People

The team

Current members first, then alumni and the collaborators whose names appear most often on our papers.

Principal investigator

  • Martin R. Prince, MD, PhD

    Professor of Radiology, Weill Cornell Medicine

    Adjunct Professor of Radiology, Columbia University Vagelos College of Physicians and Surgeons

    Research

    Contrast-enhanced MR angiography, gadolinium safety, and deep-learning volumetry on routine abdominal MRI. Full profile

Current team

Alumni

Frequent collaborators

  • Jon D. Blumenfeld, MDThe Rogosin Institute · Department of Medicine, Weill Cornell Medicine
  • Mert R. Sabuncu, PhDElectrical and Computer Engineering, Cornell University and Cornell Tech
  • Daniil Shimonov, MDThe Rogosin Institute · Department of Medicine, Weill Cornell Medicine
  • James M. Chevalier, MDThe Rogosin Institute · Department of Medicine, Weill Cornell Medicine
  • Hanna Rennert, PhDPathology and Laboratory Medicine, Weill Cornell Medicine
  • George Shih, MDDepartment of Radiology, Weill Cornell Medicine
  • Akshay Goel, MDDepartment of Radiology, Weill Cornell Medicine · Google
  • Kurt Teichman, MSDepartment of Radiology, Weill Cornell Medicine

Publications

Selected papers

Lab members are in bold. Each title links to the publisher through its DOI.

Selected list pulled from PubMed. Years follow the print issue.

    Full list on PubMed (external site)

    News

    Recent updates

    Contact

    Find us

    Department of Radiology
    Weill Cornell Medicine
    416 East 55th Street
    New York, NY 10022

    Points of contact

    • Hreedi Dev, BA Department of Radiology, Weill Cornell Medicine
    • Sophia Coraci, MS Research team
    Email addresses not added yet. Encode each with tools/encode_email.py and paste it into data-e.

    Dr. Prince online