Abstract: TH-OR023
TraceOrg 2.0: Automated Instance-Level Renal Cyst Phenotyping in ADPKD MRI
Session Information
- Artificial Intelligence in Kidney Care: From Pathology to Prediction
October 22, 2026 | Location: Room 711, Convention Center
Abstract Time: 05:10 PM - 05:20 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- He, Xinzi, Weill Cornell Medicine, New York, New York, United States
- Prince, Martin R., Weill Cornell Medicine, New York, New York, United States
- Csernus, Emoke, Weill Cornell Medicine, New York, New York, United States
- Sabuncu, Mert, Weill Cornell Medicine, New York, New York, United States
Background
Total kidney volume (TKV) is central to autosomal dominant polycystic kidney disease (ADPKD) risk assessment but does not count or measure individual cysts. TraceOrg previously automated kidney, liver, and total cyst volume measurement. Here we present TraceOrg 2.0 that extends our web platform to instance-level renal cyst phenotyping.
Methods
T2-weighted ADPKD MR images with manual cyst annotations (n=45) were evaluated using our new InstaBound cyst instance model integrated into the TraceOrg workflow. The model outputs the cyst foreground mask and an instance-aware signed-distance representation, followed by instance separation and 3D reconstruction. Quantitative biomarkers included total cyst volume (TCV), total cyst number (TCN), cyst-size distribution across volume bins, and largest cyst volume. Performance was assessed with foreground Dice, panoptic quality (PQ), intraclass correlation coefficient (ICC) for cyst biomarkers, and predicted-reference cyst-size profile matching, comparing to radiologist labels.
Results
TraceOrg 2.0 achieved median cyst foreground Dice of 0.91 and median instance-level PQ of 0.58 (segmentation quality 0.81; recognition quality 0.71). Volume-based biomarkers showed excellent agreement, including TCV (ICC 0.999; r=0.999) and TCN (ICC 0.946). Largest cyst volume showed strong agreement (ICC 0.92; r=0.951), with median absolute error of 1.8 mL. Predicted cyst-size profiles matched reference annotations across <0.1, 0.1-1, 1-10, 10-100, and >100 mL bins, with median histogram intersection of 0.94 for cyst counts and 0.95 for cyst-volume fractions.
Conclusion
TraceOrg 2.0 is a web-calculator now providing automated instance-level cyst phenotyping for ADPKD MRI in addition to TKV for quantifying cyst number, cyst-size distribution, total cyst burden, and dominant-cyst burden in a reproducible TraceOrg workflow.
Acknowledgment
NIH subcontract grant from the PRK-RRC and departmental funds.
TraceOrg 2.0 cyst phenotyping in ADPKD MRI. (A) Axial and coronal MRI crops with model-predicted cyst instances overlaid in color. (B) Cyst-size distribution by volume bin. (C) Largest-cyst volume agreement between reference and model prediction.
Funding
- NIDDK Support