Kidney Week

Abstract: SA-PO0060

Automated Quality Control System for Artificial Intelligence (AI)-Based Measurements of Total Kidney Volume

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

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

Abigail Green

• 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

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

Francesca Fati, MS

• Bozkurt, Elif G., Mayo Clinic Minnesota, Rochester, Minnesota, United States

Elif G. Bozkurt, MD

• Khalifa, Muhammed, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Muhammed Khalifa, MD

• Klug, Jason R., Mayo Clinic Minnesota, Rochester, Minnesota, United States

Jason R. Klug, PhD

• Dachowicz, Adam P., Mayo Clinic Minnesota, Rochester, Minnesota, United States

Adam P. Dachowicz, PhD

• Cook, Cole J., Mayo Clinic Minnesota, Rochester, Minnesota, United States

Cole J. Cook, MS, PhD

• Conte, Gian Marco, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Gian Marco Conte, MD, PhD

• Shi, Tubo, Mayo Clinic Minnesota, Rochester, Minnesota, United States

Tubo Shi, MBChB

• Potretzke, Theodora A., Mayo Clinic Minnesota, Rochester, Minnesota, United States

Theodora A. Potretzke, MD

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

Timothy L. Kline, MS, PhD

Background

Automated deep learning-based segmentation is increasingly used in medical imaging to enable rapid biomarker extraction. Time-consuming manual QC of segmentation outputs remains standard practice to ensure segmentation quality prior to downstream analysis. In autosomal dominant polycystic kidney disease (ADPKD), reliable segmentation is essential for accurate total kidney volume (TKV) measurement, which is a key biomarker for disease monitoring.

We developed and evaluated a 3D deep learning model that classifies input MR volumes and their automated segmentation outputs into three categories (Accept, Reject, and Rework) based on their predicted downstream success after manual review. Criteria contributing to these multiclass classification labels include image quality, image coverage of native kidneys, and segmentation accuracy.

Methods

We used a dataset of 16,492 abdominal magnetic resonance images (T2-weighted coronal scans) from patients with ADPKD for this study. Model training utilized 15,292 images, and performance was validated on a balanced set of 600 images (200 per class). A final balanced holdout test set of 600 images was used to evaluate model performance, with labels from two trained analysts and a radiologist as independent reference standards.

Results

The top-performing model (DenseNet121) achieved a median F1-score of 0.82 and median accuracy of 82% across reader-defined reference standards. Model-reader agreement measured by Cohen's κ ranged from 0.69-0.78, overlapping with the human interobserver range of 0.73-0.81. Most disagreement occurred at the subjective Accept-Rework boundary, consistent with demonstrated interobserver variability within both the validation and test datasets.

Conclusion

Downstream-aware multiclass classification models can provide scalable quality control of MR images and their automated segmentation outputs. This approach could be considered for: (i) clinical triage after automated segmentation to focus radiologist effort on flagged cases, and (ii) for automated data curation in research studies where time consuming review of images and automated segmentations is a significant bottleneck.

Acknowledgment

During preparation of this abstract, the authors used ChatGPT, OpenAI, GPT-5.4 Pro, for limited editorial suggestions related to wording and concision. No scientific content, data analysis, interpretation, images, or conclusions were generated by the tool. The authors reviewed and edited all text and take full responsibility for the integrity and accuracy of the submitted content.

Funding

• Private Foundation Support