Artificial Intelligence for Total Kidney Volume Measurements in Pediatric ADPKD: Comparison of Three-Dimensional Ultrasonography with Magnetic Resonance Imaging

Background: Identifying biomarkers to predict ADPKD progression in children is imperative. Total kidney volume (TKV), utilized in adults, holds promise as a biomarker for children as well. MRI, the standard method for measuring TKV, is often impractical for children. Recently, 3D ultrasound (US) has emerged as a low-cost, accessible alternative. This study aims to use 3D US and Artificial Intelligence (AI) deep learning techniques to measure TKV in pediatric ADPKD patients accurately. We will validate 3D US measurements by comparing them to AI MR-based TKV measurements. Methods: Using a Philips EPIQ scanner with a panoramic 3D X6-1 matrix array probe and MR scanners with 3T and 1.5T magnetic field strength, we acquired volumetric 3D-US images and MR images to segment the kidney automatically. A deep convolutional network was trained on retrospective 3D US images (n=130) of ADPKD adult subjects using a C5-1 probe and evaluated on prospectively acquired 3D US volumes (n=22). Training employed 5-fold cross-validation with images stratified by patient and kidney volume into training and validation (80/20) cohorts. We compared AI 3D US-based and MR-based TKV measurements using linear regression and Bland–Altman plots. Results: We analyzed 11 patients with genetically defined ADPKD carrying either PKD1 or PKD2 pathogenic variants. A significant correlation was observed between the reference standard total kidney volume (TKV) measured by MRI and both manual and automated 3D ultrasound (US) volumes (r=0.95, P<0.001 and r=0.86, P=0.001, respectively). The US volumes were lower compared to MR volumes, with a bias of -20.5% and -43.3% for the manual and automated measurements, respectively. Figure 1 illustrates two representative cases from the study. Conclusion: Volumetric TKV measurements in pediatric ADPKD are feasible through 3D US images. Enhanced model training, incorporating pediatric data utilizing the X6-1 probe, is anticipated to enhance segmentation performance. Funding: NIDDK Support