SwinDiff_DAS: Data Augmentation for Segmentation of the Urinary Collecting System Based on the Shift-Window Transformer and Diffusion Model

Causes of blood in the urine, such as stones and cancer, are evaluated with a dedicated CT Urography (CTU) examination. However, identifying and evaluating the urinary tract is a time-consuming and challenging process. Accurate segmentation of the urinary tract could assist radiologists in efficiently tracking these subtle targets and automatically detecting ureteral cancers and stones. Most of the current deep learning approaches that exhibit excellent performance in segmentation rely on fully supervised learning, meaning their performance is closely tied to the size and quality of the dataset used for training. However, obtaining a segmentation dataset with both large size and high quality is challenging due to the time-consuming and expertise-intensive process of manual labeling, especially for ureters. Hence, we propose the shift-window transformer-based diffusion model for synthesizing CTU images with corresponding masks of the ureters. Our approach involves two steps: unconditional mask synthesis and conditional urography synthesis. Initially, we create an unconditional diffusion model for generating masks. Subsequently, we apply the label of ureters to the conditional diffusion model for urography synthesis. During the sampling stage, we first synthesize masks and then input them into the conditional shift-window transformer-based diffusion model to generate the corresponding urography images. The segmentation results demonstrate a significant improvement when utilizing the training data that incorporates synthesized data generated by our approach.