Two-Plane AI-Assisted Renal Ultrasound for Grading Unilateral Hydronephrosis in Infants: A Single-Center Proof-of-Concept Internal Validation Study.
TL;DR
This single-center proof-of-concept internal validation study suggests that a simplified two-plane AI-assisted workflow could closely match expert consensus grading in selected infants with unilateral hydronephrosis.
Abstract
Background
Renal ultrasound is the first-line imaging method for postnatal hydronephrosis, but its severity grading remains partly reader dependent. Although previous studies have applied artificial intelligence (AI) to pediatric hydronephrosis assessment, the clinical feasibility of a simplified two-still-image workflow in a homogeneous infant cohort remains insufficiently defined.
Methods
This retrospective single-center proof-of-concept diagnostic accuracy study was conducted among 186 infants younger than 12 months with unilateral hydronephrosis who had one protocol-compatible transverse and one sagittal renal ultrasound image. The AI workflow classified hydronephrosis as mild, moderate, or severe using a two-branch convolutional neural network (CNN). Model performance was assessed using stratified patient-level five-fold internal validation. The reference standard was expert consensus based on the complete ultrasound examination. Study categories were aligned descriptively with accepted Society for Fetal Urology (SFU) and Urinary Tract Dilation (UTD) severity concepts but were not intended to replace these systems.
Results
A total of 186 infants were included. Expert consensus classified 79 cases as mild, 62 as moderate, and 45 as severe. The AI-assisted workflow showed exact agreement with expert consensus in 159 of 186 infants, corresponding to 85.5%, with a linear weighted kappa of 0.83. For clinically significant hydronephrosis, defined as moderate or severe disease, the AI workflow showed a sensitivity of 92.5%, a specificity of 86.1%, an accuracy of 89.8%, a positive predictive value of 90.0%, and a negative predictive value of 89.5%. The ordinal-score area under the curve (AUC), which is based on class labels rather than calibrated probability outputs, was 0.92 for clinically significant hydronephrosis and 0.96 for severe hydronephrosis. Because patient-level probability outputs were unavailable, formal calibration, decision curve analysis, threshold adjustment, and individualized risk estimation could not be performed.
Conclusions
This single-center proof-of-concept internal validation study suggests that a simplified two-plane AI-assisted workflow could closely match expert consensus grading in selected infants with unilateral hydronephrosis. The workflow should be considered a research-only ordinal class assignment tool. This study does not establish clinical readiness, external generalizability, calibrated risk prediction, or diagnostic superiority over routine reporting. Prospective multicenter validation with preserved probability outputs and formal calibration is needed before any clinical implementation.