An enhanced YOLOv10 architecture for high-sensitivity and high-specificity lung cancer detection

一种增强型YOLOv10架构,用于高灵敏度和高特异性肺癌检测

阅读:1

Abstract

Lung cancer detection using computed tomography (CT) imaging is a critical task for early diagnosis and improved patient outcomes. However, accurate identification of small and low-contrast pulmonary nodules remains challenging due to variations in nodule size, shape, and complex background interference. To overcome these challenges, we propose HARM-YOLO, an enhanced object detection framework based on YOLOv10, specifically designed for lung cancer detection in CT scans. Our model incorporates a multi-dimensional receptive field feature extractor (C2f-MDR), a decoupled neck architecture (DENeck), series and parallel receptive field enhancement modules (SRFEM and PRFEM), and a background attention mechanism to strengthen multi-scale feature representation and suppress irrelevant signals. Extensive experiments on the LIDC-IDRI and LUNA16 datasets demonstrate that HARM-YOLO achieves a mean average precision (mAP@0.5) of 91.3% and sensitivity of 92.7%, outperforming state-of-the-art methods including YOLOv5, ELCT-YOLO, and MSG-YOLO by significant margins. With an optimal balance of 92.7% sensitivity and 89.7% precision, our framework effectively detects true nodules while minimizing false positives, addressing key needs for computer-aided diagnosis in clinical screening. Furthermore, compared against segmentation-based approaches such as nnUNet and Swin-UNet, HARM-YOLO maintains superior performance on small nodules (≤6 mm) and real-time inference speed suitable for large-scale lung cancer screening programs. Our results highlight the potential of this YOLOv10-based object detection system as a robust and efficient tool for enhancing early lung cancer detection and supporting clinical decision-making.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。