YOLO-MF: A Multi-Branch Feature Sensing and Fine-Grained Calibration Network for Gastroenteroscopic Polyp Detection

The traditional gastroenteroscopy process relies heavily on clinicians’ empirical interpretation, suffering from issues like high missed detection rates, delayed real-time diagnostic responses, and low efficiency in large-scale screening—challenges that hinder its ability to meet the dual demands for accuracy and timeliness in early gastrointestinal cancer screening. To address this, we propose YOLO-MF (MultiFusion), a detection framework that enhances lesion sensitivity and polyp-type discrimination via a hierarchical feature interaction network. This network incorporates two novel modules: the Dual-Feature Perception Block (DFP) and Fine-Grained Feature Calibration Block (FGFC). By integrating open-source datasets with electronic cases from tertiary hospitals, we constructed a multiclass polyp dataset for model training and validation. Experimental results show YOLO-MF achieves an F1 score of 0.852 and mAP of 0.922, significantly outperforming pre-optimization baselines in complex-scenario polyp detection.