DHC-YOLO: An Improved YOLOv8 for Lesion Detection in Medical Images*

Detecting lesions in medical images is challenging due to variability in morphology, size, and boundaries across different diseases. We introduce DHC-YOLO, a novel approach that integrates multi-scale dilated attention, multi-head self-attention, and concatenation within the YOLOv8 framework. This method enhances semantic information aggregation across scales while minimizing redundancy in self-attention without adding computational costs. Our evaluation on brain tumor, colorectal polyp, and esophageal cancer datasets shows DHC-YOLO outperforms baseline YOLOv8 and other state-of-the-art models, achieving <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$mAP_{50}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$mAP_{50:95}$</tex> scores of 88.3 % and 73.5 %, 88.8 % and 67.2 %, and 52.2 % and 21.7 %, respectively. These results highlight the effectiveness of DHC-YOLO for lesion detection.