BLADDER WALL SEGMENTATION AND CANCER ASSESSMENT USING ATTENTION U-NET WITH ADVERSARIAL MECHANISM AND TWO-LAYER LEVEL SETS

Abstract

Improving the segmentation of bladder mucosa and serosa is a significant advancement for enhancing the diagnosis and treatment of bladder cancer, leading to a better understanding of the disease and improved decision-making accuracy for physicians. However, the primary aim of multi-region segmentation in bladder cancer is to facilitate the staging of the disease, specifically distinguishing between muscle-invasive bladder cancer and nonmuscle-invasive bladder cancer. This differentiation is crucial as urothelial cancers become increasingly aggressive as they infiltrate and penetrate the bladder wall. Yet, none of the existing segmentation methods have effectively addressed this critical issue. In our study, an attention U-Net with an adversarial mechanism is proposed, combining with two-layer level sets, to address these challenges. The attention U-Net with an adversarial mechanism captures the inner and outer boundaries of the bladder, including the tumor. Moreover, a two-layer level sets is proposed to further catching subpixels of the bladder wall and gets results with Dice score of 0.91. By calculating the convexity of the two-level sets, the two types of tumor are successfully classified, achieving an experimental accuracy of 0.88. This innovative approach significantly contributes to the field of bladder cancer segmentation and classification, providing valuable insights for improved diagnosis and treatment strategies.