Noninvasive Risk Stratification Based on Renal Tubular Injury Phenotypes: A Deep Learning Study for Predicting...

Area of Science:

• Nephrology and Urology
• Medical Imaging and Artificial Intelligence
• Pediatric Medicine

Background:

• Vesicoureteral reflux (VUR) can lead to kidney damage through ascending infections and inflammation.
• Current diagnostic methods like voiding cystourethrography (VCUG) are invasive, while renal scintigraphy lacks quantitative criteria.
• Accurate VUR risk stratification is crucial for timely and appropriate clinical management.

Purpose of the Study:

• To develop and validate a deep learning model for non-invasive VUR risk stratification using renal scintigraphy.
• To assess the model's ability to detect phenotypic signals of tubular dysfunction and cortical involvement.
• To evaluate the potential of the model to reduce reliance on invasive diagnostic procedures.

Main Methods:

• A deep learning model, MedSwinNet, was developed using a Swin Transformer backbone with multi-scale fusion and attention mechanisms.
• The model was trained and tested on 99mTc-DMSA scintigraphy data from 346 children with febrile urinary tract infections.
• Performance was evaluated using accuracy metrics for severe-side and bilateral-side input settings.

Main Results:

• MedSwinNet achieved high accuracies of 0.8290 (severe-side) and 0.7997 (bilateral-side) in VUR risk stratification.
• The model demonstrated stable discriminative performance and favorable generalization capabilities.
• Quality control analyses confirmed consistent image quality metrics, minimizing dataset shift bias.

Conclusions:

• Deep learning analysis of renal scintigraphy enables non-invasive quantification of VUR-associated renal involvement.
• MedSwinNet supports clinical decision-making regarding the necessity of VCUG, potentially reducing unnecessary invasive procedures.
• This approach offers improved clinical risk-stratified management for children with VUR.