| 摘 要: 针对子宫内膜癌MR图像中病灶边界模糊、形态不规则及尺度多变等问题，提出一种基于改进YOLOv8的子宫内膜癌病灶分割模型。首先，在YOLOv8的骨干网络中嵌入基于跨空间学习的高效多尺度注意力模块；其次，以焦点调制模块替代SPPF结构；最后，在颈部网络中设计融合重参数化思想与改进CSP(Cross Stage Partial)结构的双路径特征融合模块。实验结果表明，该模型在Recall、IoU、Precision和DSC上分别达到92.2%、80.1%、96%和88%，验证了所提方法在子宫内膜癌病灶分割任务中的有效性与先进性。 |
| 关键词: 深度学习 YOLOv8 子宫内膜癌 图像分割 |
|
中图分类号: TP391
文献标识码:
|
| 基金项目: 中国上海市科技创新行动计划资助项目（22S31903700）；中国上海医院发展中心联合影像联合研发计划资助项目（2022SKLY-12） |
|
| Improved YOLOv8-based Model for Endometrial Cancer Lesion Segmentation |
|
Zhang Qianqian1, Ying Jie1, Wang Yu1, Fu Le2
|
1.School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology;2.Department of Radiology, Shanghai First Maternity and Infant Hospital, Tongji University School of Medicine
|
| Abstract: To address the challenges of blurred lesion boundaries, irregular shapes, and variable scales in endometrial cancer MR images, an improved YOLOv8-based model for endometrial cancer lesion segmentation is proposed. First, a cross-space learning-based efficient multi-scale attention module is embedded in the backbone of YOLOv8. Second, the SPPF structure is replaced with a focal modulation module. Finally, a dual-path feature fusion module, integrating re-parameterization and an improved Cross Stage Partial (CSP) structure, is designed in the neck of the network. Experimental results demonstrate that the proposed model achieves Recall, IoU,Precision, and DSC scores of 92.2%, 80.1%, 96%, and 88%, respectively, validating its effectiveness and superiority in endometrial cancer lesion segmentation tasks. |
| Keywords: deep learning YOLOv8 endometrial cancer image segmentation |
