Automated detection of posterior urethral valves in voiding cystourethrography images: A novel AI-Based pipeline for enhanced diagnosis and classification

Introduction: Posterior Urethral Valves (PUV) are rare congenital anomalies of the male urinary tract that can lead to urethral obstruction and increased risk of kidney disease. Traditional diagnosis relies on subjective interpretation of imaging techniques. This study aimed to automate and increase accuracy of PUV detection in voiding cystourethrography (VCUG) images using an AI-based pipeline. The main objective was to detect presence of PUV based on urethral ratio calculated automatically from segmented urethra region.

Methods: A total of 181 VCUG images were evaluated by 9 clinicians to determine presence of PUV. Various different encoders (DenseNet, MobileNet, ResNet and VGG) were combined with Unet and Unet++ architectures to segment the urethra region. Some preprocessing and postprocessing steps were investigated to improve segmentation performance. Urethral ratios were automatically calculated with image processing and morphological operations. Finally, samples were classified between PUV or non PUV based on urethral ratio.

Results: An overall classification accuracy of 81.52 % was achieved between PUV and non PUV cases. DenseNet201 combined with Unet achieved the best overall segmentation performance (Dice score coefficient 66.15 %). Optimal cut-off value of urethral ratio for PUV detection was determined as 2.01.

Conclusion: PUV detection from VCUG images through automated segmentation and processing can reduce subjectivity and decrease physician workloads. The proposed approach can serve as a foundation for future efforts to fully automate PUV diagnosis and follow-up.