Abstract: FR-PO815
Deep Learning-Based Instance Medullary Pyramid Segmentation in Routine CT Examinations
Session Information
- Transplantation: Clinical - Outcomes
November 04, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: Transplantation
- 2002 Transplantation: Clinical
Authors
- Gregory, Adriana, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Moustafa, Amr, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Poudyal, Bhavya, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Denic, Aleksandar, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Rule, Andrew D., Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Kline, Timothy L., Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
Background
The kidney accounted for ~60% of the total transplanted organs in the U.S. in 2021. We recently demonstrated that smaller medullary volume predicted allograft loss independent of donor’s and recipient’s characteristics. In addition, the number of medullary pyramids has been suggested as a surrogate for nephron endowment, but manual quantification of pyramids is time consuming. Therefore, we developed the first fully automated approach for segmenting and differentiating individual medullary pyramids from angiogram-phase CTs.
Methods
Contrast-enhanced axial abdominal CTs were collected from 178 predonation living kidney donors. CTs from 158 subjects were used to train and validate a deep learning model that could automatically segment medullary pyramid edges and cores. 20 CT images were held out for testing. The nnU-Net framework was used to train a 5-fold cross validation ensemble model. Manual segmentations from two independent readers were used to establish interobserver variability. We used the Dice score to evaluate the segmentation similarity and Bland-Altman analysis to estimate the bias in pyramid count and volume with the reference.
Results
The test set interobserver Dice score was 0.93. The automated method had a Dice score of 0.82 compared to the first reader and 0.81 compared to the second reader. The predicted medullary pyramid count and average pyramid volume bias±sd with the reference standard was 2±5 pyramids and -0.02±0.3 ml, respectively.
Conclusion
A fully automated instance medullary pyramid segmentation method was developed and tested for healthy kidney CT scans. This approach unlocks the potential of medullary size and count as a kidney biomarker. Further work needs to determine whether size and number of medullary pyramids relates to nephron size and number and associates with kidney disease risk factors and outcomes.
Funding
- NIDDK Support