Abstract: TH-PO466
Automated Segmentation of Individual Kidney Cysts in Routine Abdominal CT Images
Session Information
- Cystic Kidney Diseases: Clinical Assessment and Therapeutic Directions
October 24, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Genetic Diseases of the Kidneys
- 1201 Genetic Diseases of the Kidneys: Cystic
Authors
- Gregory, Adriana, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Im, Jeeho, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Khalifa, Muhammed, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Chebib, Fouad T., Mayo Clinic in Florida, Jacksonville, Florida, United States
- Torres, Vicente E., Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Harris, Peter C., Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Erickson, Bradley J., 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
Autosomal dominant polycystic kidney disease (ADPKD) is a genetic disorder characterized by the development and growth of many cysts in the kidneys. Total kidney volume (TKV) is the main imaging biomarker used for disease progression prediction; however, advanced biomarkers like total cyst number (TCN) and cyst-parenchyma surface area have been shown to provide complementary information to TKV in MR imaging. Automated methods to segment individual kidney cysts in CT images of patients affected by PKD are still needed.
Methods
A 3D deep learning model using the nnUNet architecture was trained (n=60 CT images) to learn cyst edges and cores. Then, a postprocessing algorithm was applied to convert the semantic edge-core representation to instance (individual) cyst level segmentations. A 5-fold cross-validation was done to improve model robustness. The model was tested on 15 CT images not seen during training. The test set was manually segmented by two blinded readers to establish interobserver variability and the target performance for the model. The similarity between segmentations was assessed by Dice score, Bland Altman, and linear regression.
Results
The automated segmentation approach performed on par with human readers. Table1 shows the assessment metrics results. Significant agreement was observed between the readers and between each reader with the automated model. Figure1 presents an example test case.
Conclusion
This is the first model trained to segment individual cysts in contrast-enhanced CT images of patients affected by ADPKD. The initial results show good agreement between the model and segmentations of two expert readers.
Similarity metrics
| Reader1 vs Reader2 | Reader1 vs Auto | Reader2 vs Auto | |
| †Dice Score | 0.88 ± 0.07 | 0.86 ± 0.05 | 0.88 ± 0.06 |
| TCV Bias% [95% CI] | 15.1 [9 to 22] | 15.1 [9 to 21] | 0.00 [-8 to 8] |
| TCV correlation (r) | 1.00 | 1.00 | 1.00 |
| TCN correlation (r) | 0.97 | 0.97 | 0.96 |
†Mean ± st. dev. TCV and TCN: total cyst volume and number
Funding
- NIDDK Support