Scholarlink

<p>Due to inherent limitations in identifying small-scale targets, accurately detecting lung nodules in computed tomography (CT) scans remains challenging, often leading to missed diagnoses and reduced diagnostic specificity. This study proposes YOLOv8 EMFA, an enhanced deep learning framework that addresses these key limitations through three key architectural innovations. Firstly, we redesigned the feature fusion hierarchy by integrating improved Neck components from Gold YOLO’s cross stage aggregation mechanism, significantly enhancing the flow of contextual information. Secondly, the fused ScalSeq module implements cascaded multi receptive field operations in the feature pyramid network, effectively solving the scale variance challenge in small nodule representation. Thirdly, our dual branch attention architecture combines the adaptive spatial feature (ASF) module for foreground and background discrimination with a novel C2f EMFA block that utilizes multi-path hierarchical feature extraction, achieving the optimal balance between localization accuracy and computational efficiency. Through comprehensive validation on the LUNA16 benchmark dataset, the YOLOv8-EMFA framework proposed in this study demonstrates excellent performance, with an average precision (AP@0.5 )Reached 90.7%, which is a significant improvement of 2.8% points compared to the baseline YOLOv8 model, and 1.9 and 4.4% points higher than the YOLOv8-GoldYOLO and CAF-YOLO models, respectively, outperforming multiple existing small object detection models. These technological advancements make YOLOv8-EMFA a powerful multiscale lung nodule detection framework with potential significance for early diagnosis and longitudinal treatment monitoring of lung cancer.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

YOLOv8-EMFA: an enhanced multi-feature aggregation network for pulmonary nodule detection in CT scans

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Mingze Sun,
  • Peixue Liu,
  • Xinyue Han,
  • Ziyi Zhu,
  • Xinchan Li,
  • Han Zhang

摘要

Due to inherent limitations in identifying small-scale targets, accurately detecting lung nodules in computed tomography (CT) scans remains challenging, often leading to missed diagnoses and reduced diagnostic specificity. This study proposes YOLOv8 EMFA, an enhanced deep learning framework that addresses these key limitations through three key architectural innovations. Firstly, we redesigned the feature fusion hierarchy by integrating improved Neck components from Gold YOLO’s cross stage aggregation mechanism, significantly enhancing the flow of contextual information. Secondly, the fused ScalSeq module implements cascaded multi receptive field operations in the feature pyramid network, effectively solving the scale variance challenge in small nodule representation. Thirdly, our dual branch attention architecture combines the adaptive spatial feature (ASF) module for foreground and background discrimination with a novel C2f EMFA block that utilizes multi-path hierarchical feature extraction, achieving the optimal balance between localization accuracy and computational efficiency. Through comprehensive validation on the LUNA16 benchmark dataset, the YOLOv8-EMFA framework proposed in this study demonstrates excellent performance, with an average precision (AP@0.5 )Reached 90.7%, which is a significant improvement of 2.8% points compared to the baseline YOLOv8 model, and 1.9 and 4.4% points higher than the YOLOv8-GoldYOLO and CAF-YOLO models, respectively, outperforming multiple existing small object detection models. These technological advancements make YOLOv8-EMFA a powerful multiscale lung nodule detection framework with potential significance for early diagnosis and longitudinal treatment monitoring of lung cancer.