Artificial Intelligence for Pediatric Vesicoureteral Reflux Assessment on Voiding Cystourethrography: A Scoping Evidence Map of Diagnostic Performance and Clinician-Assisted Interpretation

Technical Summary: Artificial Intelligence for Pediatric Vesicoureteral Reflux Assessment on Voiding Cystourethrography

Problem Statement
Vesicoureteral reflux (VUR) is a significant pediatric urologic condition where retrograde urine flow from the bladder to the ureter and renal collecting system necessitates accurate diagnosis and grading. The standard diagnostic method, Voiding Cystourethrography (VCUG), relies on the International Reflux Grading System (Grades I–V). However, visual grading is inherently subjective, particularly at the Grade III–IV boundary. This variability influences critical clinical decisions regarding antimicrobial prophylaxis, surgical intervention (e.g., endoscopic injection or ureteral reimplantation), and long-term renal risk assessment. While Artificial Intelligence (AI) and deep learning have been applied to automate or assist in VCUG interpretation, the literature is characterized by heterogeneity in target tasks, imaging inputs, architectures, and validation strategies. Consequently, it remains unclear whether current AI models offer autonomous superiority over radiologists or if they serve primarily as tools to improve grading consistency and reduce interobserver variability.

Methodology
This study employs a scoping evidence map design rather than a formal diagnostic meta-analysis, adhering to the PRISMA Extension for Scoping Reviews (PRISMA-ScR). The synthesis was conducted using a computer-assisted workflow based on two exported datasets generated via the Elicit tool:

1. A screening dataset containing 500 records.
2. An extraction dataset containing detailed data from 13 included studies published between 2019 and 2025.

Inclusion Criteria:

• Pediatric patients or pediatric imaging data.
• Evaluation of AI, machine learning, deep learning, or computer-aided diagnostic approaches.
• Focus on VUR detection, grading, low/high-grade classification, AI-assisted interpretation, or clinically related VUR prediction using VCUG or related data.
• Reporting of diagnostic performance, reliability, agreement, or clinician-comparison metrics.

Exclusion Criteria:

• Non-pediatric studies, non-human data, editorials, isolated case reports, or technical reports lacking clinically interpretable outcomes.

Classification Strategy:
To address methodological heterogeneity, studies were categorized into four groups:

1. Direct VCUG-based VUR grading: Models attempting to assign standard grades.
2. Low/High-grade classification: Binary or severity-stratified models.
3. AI-assisted clinician interpretation: Studies evaluating AI's impact on human reader performance.
4. Indirect clinical prediction: Models predicting outcomes (e.g., recurrent UTI) rather than grading reflux directly.

Due to the lack of independent duplicate screening and full-text verification in the source exports, the synthesis is explicitly framed as an evidence map, prioritizing narrative synthesis and tabulation over pooled statistical analysis.

Key Contributions and Results
The review identified 13 studies, revealing a landscape of promising but heterogeneous AI performance:

• Diagnostic Performance:

  • The strongest externally validated evidence comes from Li et al. (2024), whose Deep-VCUG model achieved external Area Under the Curve (AUC) values of 0.944 for unilateral and 0.924 for bilateral reflux.
  • Khondker et al. (2022) reported an AUC of 0.84 and demonstrated a 3.6-fold improvement in grading reliability compared to traditional assessment.
  • Wu et al. (2025) showed substantial improvements in clinician performance when supported by AI, with clinician AUCs rising from ~0.63–0.73 to ~0.95–0.96 for side-specific reflux.
  • Several smaller studies reported near-perfect metrics (e.g., accuracy approaching 100%, AUC near 1.0), but the review cautions these findings due to risks of overfitting, small sample sizes, and limited external validation.

• Clinician Comparison and Assistance:

  • Direct comparisons between AI and radiologists are limited. Most studies used radiologist labels as reference standards without independently quantifying radiologist baseline performance.
  • However, studies focusing on AI-assisted interpretation indicate that AI can significantly reduce inter-rater variability and improve the performance of both junior and senior clinicians, particularly in recognizing high-grade VUR.

• Evidence Quality and Gaps:

  • Indirect vs. Direct: The review distinguishes between direct grading (most clinically relevant) and indirect prediction models (e.g., predicting recurrent UTI risk), noting that the latter do not prove AI's ability to grade reflux severity.
  • Management Impact: No included study definitively demonstrated that AI changes clinical management decisions (e.g., altering antibiotic prophylaxis or surgical thresholds) or improves patient outcomes.
  • Limitations: The evidence base suffers from a lack of standardized reporting (sensitivity, calibration, decision-curve analysis), limited external validation across diverse datasets, and a scarcity of prospective, blinded reader studies comparing AI-alone, clinician-alone, and AI-assisted arms.

Significance and Claims
The paper concludes that while AI algorithms show promising diagnostic performance for pediatric VUR assessment on VCUG, the current evidence does not support the autonomous replacement of radiologist interpretation.

The primary significance of the findings lies in the potential for AI to function as a decision-support tool rather than a standalone diagnostician. Specifically, AI may:

1. Improve grading consistency, particularly around the critical Grade III–IV boundary.
2. Reduce interobserver variability.
3. Support less experienced readers and standardize high-grade VUR recognition.

The authors emphasize that for clinical implementation, future systems must undergo prospective, multicenter, blinded reader studies that evaluate AI's impact on management decisions and patient outcomes, adhering to reporting standards such as CLAIM, TRIPOD+AI, and STARD-AI. Until such evidence is established, AI should be viewed as a tool to assist, not replace, expert clinician interpretation.