One-Stage Attention-Centric Instance Segmentation Framework for Gastrointestinal Disease Localization in Endoscopic Images

Accurate instance segmentation of gastrointestinal (GI) diseases in endoscopic images remains challenging due to visual ambiguity, diffuse lesion boundaries, and class imbalance. This study presents a one-stage, attentioncentric instance segmentation framework for the automated localization and classification of gastrointestinal diseases and anatomical landmarks in endoscopic imaging. The model employs residual feature aggregation and attention-enhanced representations and is trained using transfer learning with extensive preprocessing and data augmentation to improve generalization. Experimental results demonstrate stable convergence and strong performance, achieving an overall mAP@0.5 of 0.856, with precision and recall exceeding 0.80 for most classes and a macro-averaged F1 score of 0.83 at an optimal confidence threshold of approximately 0.49. High accuracy is observed for visually distinct anatomical landmarks and localized lesions, while inflammatory conditions exhibit reduced performance under stricter localization criteria (mAP@0.5 -$0.95 \approx 0.65-0.70$), reflecting challenges associated with diffuse pathology. These findings indicate that performance limitations are primarily data-driven rather than architectural, highlighting the potential of the proposed framework for real-time endoscopic decision support and the importance of improved data curation for clinical deployment.