BlaVeS: A Novel Hand-Labeled Dataset for Improved Bladder Vessel Segmentation with Modified U-Net

Orientation and reliable localization of abnormalities in the urinary bladder are essential for providing optimal patient-specific treatment in uro-oncology. Blood vessels serve as anatomical landmarks, assisting in orientation during cystoscopic procedures despite challenging intraoperative conditions, including variable lighting and continuous tissue deformation. Despite the importance of bladder vessel segmentation, research in this area remains limited. This work introduces the first publicly available dataset of endoscopic bladder images dedicated to vessel segmentation, capturing various scenarios faced in cystectomy. The blood vessels were manually annotated, providing a high-quality ground truth for training deep learning models. Standard architectures for vessel segmentation from the well-studied retina domain fail to achieve satisfactory results on this dataset. To address this, we propose a modified U-Net architecture incorporating deep skip connections and attention mechanisms to address the challenges posed by variable imaging conditions. These adaptions improve segmentation performance, achieving an accuracy of 0.93 and a precision of 0.55. Our dataset and method lay the foundation for advancing automated bladder vessel segmentation, facilitating improved diagnostic and therapeutic outcomes in urological imaging.