Noninvasive Risk Stratification Based on Renal Tubular Injury Phenotypes: A Deep Learning Study for Predicting Vesicoureteral Reflux in Children

Vesicoureteral reflux (VUR) can cause retrograde urine flow under voiding pressure, facilitating ascending bacterial colonisation and recurrent inflammatory responses. These processes trigger a cascade of cellular and molecular events-innate immune activation, pro-inflammatory cytokine release, oxidative stress, apoptosis and extracellular matrix deposition-thereby promoting tubulointerstitial remodelling and increasing the risk of renal parenchymal injury and scarring. Static renal 99mTc-DMSA scintigraphy primarily reflects tracer uptake by proximal tubular cells in the renal cortex and can serve as an integrated phenotypic readout of tubular dysfunction and focal cortical involvement. However, its clinical interpretation remains experience-dependent and lacks reproducible quantitative criteria, while voiding cystourethrography (VCUG), the diagnostic and grading gold standard for VUR, is limited by invasiveness and procedural burden. In this study, we collected DMSA data from 346 children with febrile urinary tract infection treated at the Second Affiliated Hospital of Wenzhou Medical University between January 2019 and January 2023 and developed a deep learning model (MedSwinNet) for VUR risk stratification. Built on a Swin Transformer backbone and enhanced with multi-scale representation fusion, a convolutional block attention module and a gated selection strategy, MedSwinNet was designed to sensitively capture phenotypic signals such as reduced proximal tubular uptake and focal cortical defects while improving robustness. On the test set, the model achieved accuracies of 0.8290 under the severe-side input setting and 0.7997 under the bilateral-side input setting, demonstrating stable discriminative performance and favourable generalisation. Quality control analyses indicated broadly consistent distributions of key image quality metrics across data splits, mitigating potential bias from dataset shift. Collectively, deep learning-based decoding of tubular dysfunction-related phenotypic readouts enables noninvasive quantification of VUR-associated renal involvement, supports decision-making on whether VCUG is warranted and may reduce unnecessary invasive procedures while improving clinical risk-stratified management.