Publication

Abstract : Colonoscopy plays a significant role in computer-aided diagnostic systems, contributing to the early detection of colorectal anomalies such as cancers and polyps. RGB-D cameras, which provides complimentary depth information represent a significant innovation in enhancing the detection and diagnosis of anomalies in colonoscopy. This study investigates the different fusion strategies of RGB and depth features to improve the semantic segmentation of polyps using the UNet framework. The strategies include using single mode RGB and depth information, RGB-D as 4-channel information, fusing depth features extracted at each level on the encoder side stages, the decoder side stages and through encoder-Decoder hybrid fusion. Experiments conducted on publicly available colonoscopy dataset show that the fusion of deep RGB and depth improves segmentation performance over single-modal approaches. Quantitative results highlight the improvements in standard segmentation metrics, while qualitative analysis underscores the ability to segment polyps of different size and shape accurately. The study highlights the potential of RGB-depth fusion as a promising approach for advancing automated polyp segmentation in real-world colonoscopy applications.