<p>Gastrointestinal endoscopy is indispensable for early detection and treatment of gastrointestinal diseases. However, the diagnostic precision and sensitivity of current computer-aided diagnosis systems remain limited, particularly for lesions with small size, low visibility, or poorly defined boundaries. In this paper, we propose a Gastrointestinal Abnormality Detection YOLO model (termed GAD-YOLO), which is built upon the YOLOv8s architecture and enhanced with multi-scale channel attention and residual fusion to improve the detection accuracy and robustness in gastrointestinal endoscopy. In the backbone, two improved modules, MSC2f and D4f, are integrated to strengthen cross-scale feature representation and hierarchical context fusion, enhancing the model’s capability in capturing multi-scale lesion features. In particular, the MSC2f module is designed to enhance the detection sensitivity for diminutive and flat lesions, which are frequently overlooked during routine gastrointestinal endoscopic examinations but are critical for early-stage disease identification. In the neck, a newly designed SEC2f module adaptively recalibrates channel responses after multi-scale feature fusion, thereby refining fused representations and reducing missed detections. The proposed model was trained and evaluated on the public Kvasir-Capsule dataset and compared with five YOLO variants, SSD, and Faster R-CNN. GAD-YOLO achieved a mean Average Precision (mAP) of 0.974, outperforming all baseline models by 1.4%–29.7%. These findings indicate that GAD-YOLO has the potential to support reliable identification of subtle and clinically relevant gastrointestinal lesions in complex endoscopic scenarios.</p> Graphical Abstract <p></p>