Small gastric polyp detection based on the improved YOLOv5

Small target polyps are prone to missed detection due to their small coverage area and little information. To address this issue, a modified PATM-YOLO polyp detection model based on YOLOv5 is proposed. The model first addresses the issue of missed detection of small polyps by constructing a detection head for identifying small polyps and using an improved Phase-Aware Token Mixing Module(PATM) attention module to increase the network’s attention to small polyps and suppress the model’s focus on non-polyp regions. Secondly, an improved Adaptively Spatial Feature Fusion(ASFF) module is proposed to fully utilize multi-scale information, enhancing the network’s feature richness. Finally, by introducing the Swin Transformer into the network and determining its optimal placement through experiments, the detection accuracy is maximized without affecting the network’s performance. After experimental comparison on the constructed dataset and the public dataset SUN, the proposed PATM-YOLO network model alleviated missed detection in dense and small polyp images, and achieved an precision of 91.3 % , which is 8.5 % higher than the baseline YOLOv5 network model. This indicates that the detection performance of this model outperforms other classical target detection networks and the original network.