Posterior Urethral Valves Classification

This study developed a multi-instance deep learning model to automatically distinguish children with posterior urethral valves (PUV) from those with mild unilateral hydronephrosis using routine renal ultrasound images. Using over 6,000 sagittal and transverse images from 157 patients, the multi-instance approach significantly outperformed single-image models, achieving an AUC of 0.961 with high specificity (0.986) and strong overall classification accuracy. Activation mapping demonstrated that the model identified clinically meaningful anatomical regions, supporting its potential as a reliable and automated diagnostic tool for pediatric PUV.