Early detection and accurate segmentation of polyps are essential for reducing colorectal cancer mortality. However, manual analysis of colonoscopy images is error-prone due to variability in polyp appearance, limited contrast, and operator fatigue. To address these challenges, we propose a deep learning-based framework that integrates object detection and semantic segmentation for precise polyp analysis. The method employs YOLOv8 to localize polyps across diverse colonoscopic imaging modalities, followed by a modified U-Net that performs fine-grained segmentation guided by attention masks derived from the detection output. This attention mechanism directs the segmentation network to focus on polyp-relevant regions. In addition, a weighted Binary Cross-Entropy loss is applied to emphasize tumor pixels during training. Experiments on a multi-modal colonoscopy dataset demonstrate that the proposed framework achieves superior performance in Dice score, Intersection over Union, and pixel-level precision and recall. These results highlight the system’s robustness and clinical relevance for enhancing polyp detection and delineation during colonoscopy. The source code is made publicly available for reproducibility at: https://github.com/greenredjr/Colon-Tumor-Detection-and-Segmetation .