BWS-Net: An Optimal Deep Learning Architecture for the Anterior Bladder Wall Segmentation using Ultrasound Imaging.

Abstract

Urodynamic tests are used to assess bladder function by measuring detrusor pressure, which requires invasive catheterization. Ultrasound bladder vibrometry offers a non-invasive approach to evaluate bladder compliance for diagnostic purposes by estimating detrusor pressure from ultrasound-generated Lamb waves in the bladder wall. Therefore, it is crucial to precisely segment the anterior portion of bladder wall prior to these assessments. Traditional segmentation methods are time-consuming and require manual intervention, while deep learning offers a promising alternative. Existing deep learning approaches are limited and often lack clinical validation. Therefore, we propose a novel deep learning network for precise segmentation of anterior bladder wall. It comprises blueprint separable convolutional layers in an encoder-decoder structure with adaptive attention-based skip connections. Performance evaluation on 8592 distinct images acquired from 64 patients using 5-fold cross-validation demonstrates that it achieves a mean Dice score and sensitivity of 0.82 and 0.85 respectively, along with a mean root mean square error of $0.67 \pm 0.35mm$ between the thickness of predicted and ground truth portions of the bladder wall. Blueprint separable convolutions and adaptive attention-skip connections improve segmentation performance with fewer computations compared to respective standard counterparts. A comparative analysis demonstrates improvements between 2-20% in the dice score with respect to some but one of the existing networks and reductions in computational complexity by $94-96\%$ with respect to all existing networks analyzed in this work. Therefore, the proposed method can be effective for accurate bladder wall segmentation and demonstrates potential for real-time application in clinical settings.