摘要
目的:提出一种基于改进YOLOv5s的免疫组化阳性细胞计数方法。方法:首先,针对阳性细胞的小目标特征,增加小目标检测层细化特征提取;其次,在颈部网络中使用双向加权特征金字塔网络替换PANet结构,实现多尺度特征融合;再次,增加坐标注意力机制CA模块,使模型更加关注小目标特征;最后,用EIoU损失函数替换原有的GIoU,增强模型检测性能。结果:在自建的免疫组化图像数据集上进行训练,改进后的模型平均精确率为89.3%,较原模型提升4.0%,且优于其他主流目标检测模型。同时,基于该方法构建的5年生存预测模型平均准确率为76.8%,平均AUC为0.81,表明模型具有较好的预测能力。结论:本研究模型能够快速检测免疫组化阳性细胞数量并有效地辅助医生进行生存预测工作。
Objective To propose a novel method for immunohistochemically positive cell counting based on the improved YOLOv5s.Methods Regarding the small target characteristics of positive cells,a small target detection layer was added to refine feature extraction.Then,a bidirectional weighted feature pyramid network was used to replace path aggregation network(PANet)in the neck network for realizing multi-scale feature fusion.Additionally,the method used coordinate attention mechanism to make the model pay more attention to small target characteristics,and replaced the original GIoU with EIoU loss function for enhancing the detection performance.Results The model was trained on the self-built immunohistochemical image dataset.The average accuracy of the improved model was 89.3%,which was 4.0%higher than the original model and surpassed mainstream target detection models.The 5-year survival prediction model constructed with the method achieved an average accuracy of 76.8%and an average area under the curve of 0.81,demonstrating its superior prediction ability.Conclusion The proposed model can quickly detect the number of immunohistochemically positive cells and effectively assist doctors in survival prediction.
作者
陈星月
贾子彦
李青
张大川
潘玲佼
沈大伟
CHEN Xingyue;JIA Ziyan;LI Qing;ZHANG Dachuan;PAN Lingjiao;SHEN Dawei(School of Electrical and Information Engineering,Jiangsu University of Technology,Changzhou 213001,China;Department of Pathology,Changzhou First People's Hospital,Changzhou 213004,China)
出处
《中国医学物理学杂志》
2025年第2期167-174,共8页
Chinese Journal of Medical Physics
基金
国家自然科学基金(62001196)
江苏省第六期333高层次人才培养工程(2022-2)
常州市5G+工业互联网融合应用重点实验室(CM20223015)
常州市应用基础研究(CJ20220064,CJ20220059)。
作者简介
陈星月,硕士研究生,研究方向:医学图像处理、目标检测,E-mail:643395909@qq.com;通信作者:贾子彦,副教授,研究方向:可见光通信、机器视觉、5G研究,E-mail:jiaziyan@jsut.edu.cn。
