Dual-generative synthesis framework: enhancing polyp segmentation in colonoscopy via mask-conditional GANs.

Background and Objective: Colorectal cancer screening through colonoscopy relies on the segmentation of polyps, which becomes difficult when polyps are small in size and flat in shape, with subtle edges, poor contrast, and large variability in appearance. Although new deep learning techniques have improved segmentation performance, most techniques have focused on architectural innovations, leaving aside the issue of limited diversity in the training set. Although generative methods have improved, they often suffer from the lack of control over structure, misalignment of masks and images, and increased computation, making them inefficient in the case of small polyps, where they are most relevant. This study aims to present a dual generative synthesis framework for data-centric augmentation to improve segmentation performance. Methods: The proposed framework divides the problem of structure and appearance into two steps. First, procedural generation is used to create realistic masks of small, flat polyps. Second, a mask-conditioned GAN generates colonoscopy images that match the generated masks in texture and lighting conditions. This process produces anatomically realistic and perfectly aligned mask-image pairs. The generated data are incorporated into a U-Net-based segmentation model and evaluated on public polyp segmentation datasets with extensive ablation studies. Results: The proposed method achieved a Dice score of 0.8786 and an Intersection over Union (IoU) of 0.7835. The model also obtained a precision of 0.8930 and a recall of 0.8648, which are significantly higher than those of the baseline U-Net model. Conclusions: The dual generative synthesis framework improves segmentation robustness by generating realistic and aligned training data for small and flat polyps. The results demonstrate its potential to enhance the reliability of automated polyp segmentation in colonoscopy images.