Comparison of Artificial Intelligence (AI)-Driven Method with Semiautomated Segmentation for Total Kidney Volume (TKV) Estimation in ADPKD

Background: ADPKD is a genetic disease characterized by progressive cyst growth, leading to renal failure. The imaging-based TKV has been accepted as prognostic biomarker in ADPKD clinical trials by the FDA. Currently, manual or semi-automated segmentation by an expert radiologist is used to measure TKV. Recent advances in AI-driven methods allow for disease progression monitoring with high repeatability and reproducibility. Objectives of this study were 1) To compare AI-driven kidney segmentation to a semi-automated expert radiology reader for TKV measurement, 2) To assess the longitudinal TKV change between methods. Methods: A retrospective cohort of MRI scans, consisting of coronal T2w and T1w images, from ADPKD subjects (n = 12 with 2 timepoints each) from a sample of CRISP and HALT PKD datasets (NIDDK Central Repository) were analyzed. One expert radiologist assessed TKV using T2w images, while visually referencing T1w images. A Deep Learning 3D U-Net model was previously developed and trained using radiologist-based kidney segmentations from 67 timepoints using ADPKD T2w images from the same CRISP dataset. The model was then applied to this separate test cohort to generate AI-driven TKV assessments. Results: Radiology and AI-driven TKV assessments demonstrated high correlation (R2 = 0.94) with a mean difference of -4.7cc (-1.5%). Figure 1A) Alignment between AI method to radiologist review. 1B) Bland-Altman plot shows similarity between TKV assessment methods.Longitudinal TKV assessments demonstrate alignment in absolute change between timepoints, radiologist (196.8cc, 13.3%) vs AI-driven (174.7cc, 11.5%). Conclusion: The proposed AI-driven approach demonstrates consistent performance to an expert radiologist and is suitable for longitudinal TKV monitoring. Despite cohort data originating from different datasets, the AI-driven approach exhibits a robust analysis feasible of assessing TKV from images acquired on multiple sites and scanners. Future work is to further refine the AI-driven model with additional training and testing data and determine reader variability for each analysis method. Funding: NIDDK Support