Multi-institutional Validation of Improved Vesicoureteral Reflux Assessment With Simple and Machine Learning...

Area of Science:

• Urology
• Medical Imaging
• Artificial Intelligence

Background:

• Vesicoureteral reflux (VUR) grading from voiding cystourethrograms (VCUGs) is subjective and lacks reliability.
• Current grading methods exhibit low inter- and intra-rater agreement, impacting clinical decision-making.

Purpose of the Study:

• To enhance the reliability of VUR grading using both traditional and machine learning (ML) approaches.
• To assess the effectiveness of ureteral tortuosity and dilatation features in VCUGs for VUR grading.

Main Methods:

• Collected VCUGs from a large pediatric cohort for training and external validation.
• Assessed VUR grade using 4 features: ureteral tortuosity, proximal, distal, and maximum ureteral dilatation.
• Developed and validated a ML model (qVUR) to predict VUR grade based on these features, measuring performance with AUROC.

Main Results:

• Internal inter-rater reliability for VUR grading was low (0.44), with a median agreement of 0.71.
• The ML model, qVUR, achieved an accuracy of 0.62 (AUROC=0.84) and showed stable performance across datasets.
• qVUR significantly improved VUR grading reliability by 3.6-fold compared to traditional methods (P < .001).

Conclusions:

• Machine learning-based VUR assessment offers improved reliability over current grading methods in a large pediatric population.
• The qVUR model is generalizable and robust, demonstrating comparable accuracy to clinicians.
• Further research is warranted to explore the prognostic value of quantitative measures provided by qVUR.