Discrete Diffusion Propagated Transformer For Flexible Ureteroscopic Semantic Segmentation

Ureteroscopic surgery inserts flexible ureteroscopes into the bladder and ureter to reach the kidneys, and observes renal diseases such as stones and tumors. Renal structures (e.g., pelvis, calyx, and papilla) and stones are important to locate the ureteroscope in the kidneys. This work proposes a new small deep learning framework of discrete diffusion propagated transformer for flexible ureteroscopic video semantic segmentation. Specifically, this framework first employs transformer, recombination, and fusion modules to coarsely extract structure, stone, and background regions. Moreover, a discrete diffusion propagation mechanism is introduced to optimize and refine the coarse segmentation. We evaluate our method on surgical ureteroscopic videos acquired from the operating room. The experimental results show that our proposed semantic segmentation method significantly outperforms state-of-the-art approaches. Particularly, it can improve the average dice coefficient and intersection over union from (84.7%, 73.5%) to (90.3%, 82.4%), respectively.