Kidney Week

Abstract: SA-PO0063

Deriving Kidney Function from Medical Images for Patients Affected by Polycystic Kidney Disease Using Multitask Deep Learning

Session Information

• Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
  October 24, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Dhar, Mrinal Kanti, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Mrinal Kanti Dhar, PhD

• Gregory, Adriana, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Adriana Gregory, MS

• Buchl, Samuel, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Samuel Buchl

• Holmes, Heather L., Mayo Clinic Minnesota, Rochester, Minnesota, United States

Heather L. Holmes

• Green, Abigail, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Abigail Green

• Fati, Francesca, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Francesca Fati, MS

• Kline, Timothy L., Mayo Clinic Minnesota, Rochester, Minnesota, United States

Timothy L. Kline, MS, PhD

Background

Clinically significant chronic kidney disease (CKD) is typically identified using serum creatinine-derived estimated glomerular filtration rate (eGFR), not imaging. This project investigates whether MR images contain sufficient imaging information to distinguish between patients with relatively preserved kidney function and those with reduced renal function.

Methods

Deep learning models were trained on 2300 MRIs and evaluated on an independent test set of 576 MRIs (classes 0/1: 347/229). Class labels were defined using eGFR: class 0 (eGFR ≥60 mL/min/1.73 m²) and class 1 (eGFR<60). This threshold was chosen to separate healthy to milder disease from moderate to advanced functional impairment. Four strategies were compared against a ResNet18 baseline: (1) multi-task learning (regression & classification); (2) multi-modal multi-task learning by adding age and/or sex; (3) multi-image learning combining whole-image and kidney ROI embeddings; and (4) ensemble network (ResNet18 & VGG16).

Results

The baseline model achieved 0.65 accuracy and 0.76 AUC. Among multi-task models, the two-head joint regression-classification model performed best, achieving 0.81 accuracy, 0.77 F1 score, and 0.89 AUC. In the multi-modal setting, adding sex improved performance (0.79 accuracy, 0.76 F1, 0.87 AUC), while age alone showed low sensitivity (0.31). Combining whole-image and kidney ROI features outperformed ROI-only images (0.71 vs. 0.66 accuracy; 0.77 vs. 0.71 AUC). The ensemble network achieved the highest overall performance (accuracy 0.83, AUC 0.92, sensitivity 0.88, and F1 score 0.81).

Conclusion

An MR-based CKD classifier could enable noninvasive, opportunistic identification of clinically significant CKD from routinely acquired imaging and reduce dependence on contemporaneous serum creatinine testing for screening or triage.

Funding

• Private Foundation Support