摘要
目的宫颈癌(cervical cancer,CCA)是危害女性生命健康的高发癌症之一,针对宫颈液基细胞学检查人工阅片费时费力、主观性强且医疗资源不足的问题,提出了一种基于深度学习的宫颈细胞全视野数字切片(whole slide image,WSI)自动分析方法,以提高病理医师的筛查和诊断效率。方法收集山东省立第三医院病理科2834张宫颈细胞数字切片,划分为patch数据集和WSI数据集。将patch数据集用于深度学习模型搭建、训练和测试,为了消除WSI制片和扫描过程中颜色差异的影响,使用大量数据增强方法扩增数据集以提高模型泛化能力,在YOLOv8模型基础上引入高效多尺度注意力(efficient multi-scale attention,EMA)增强模型对复杂样本的检测能力,采用精准率、召回率、全类平均精度(mAP)等指标评价模型性能。对于WSI数据集,使用TopK平均置信度方法集成检测结果来预测WSI阳性概率,使用敏感度、特异度等指标评价分类方法性能。结果Patch数据集验证集和测试集的mAP@0.5指标分别为0.717和0.652,消融实验中模型添加数据增强和注意力方法后mAP@0.5分别提升2.5%和1.3%,最终精准率为0.661、召回率为0.69、mAP@0.5为0.717;WSI数据集切片分类的敏感度为0.974、特异度为0.39。结论本研究所提方法具有良好的异常宫颈细胞检测能力,在临床上可帮助病理医师进行病灶区域的快速筛查,同时具有一定的WSI阴性阳性分类能力,可以在疾病早期筛查中进行分流管理。
Objective Cervical cancer(CCA)is one of the high-risk cancers that endanger women’s health.Aiming at the problems of time-consuming,labor-intensive,strong subjectivity,and insufficient medical resources in the manual reading of cervical fluid-based cytology,we propose an automatic analysis method for whole slide image(WSI)of cervical cells based on deep learning,which can improve pathologist’s accuracy and efficiency.Methods A total of 2834 WSI of cervical cells from Department of Pathology,Shandong Provincial Third Hospital were collected and divided into patch dataset and WSI dataset.The patch dataset was used for building,training,and testing deep learning models.To eliminate the influence of color differences during WSI production and scanning,a large number of data enhancement methods were used to expand the dataset to improve the model generalization ability.Based on the YOLOv8 model,efficient multi-scale attention(EMA)was introduced to enhance the detection ability of complex objects.The performance of the model was evaluated by precision,recall,and mean average precision(mAP).For the WSI dataset,the TopK average confidence method was used to integrate detection results to predict the positive probability of WSI,and sensitivity and specificity were used to evaluate the performance of classification methods.Results The mAP@0.5 of the patch dataset validation set and test set were 0.717 and 0.652,respectively.In the ablation experiment,after using data augmentations and adding EMA mechanisms to the model,mAP@0.5 increased by 2.5%and 1.3%respectively,and the final accuracy was 0.661,the recall was 0.69,and mAP@0.5 was 0.717 of the patch dataset.The sensitivity and specificity of the WSI classification were 0.974 and 0.39 respectively.Conclusions The method can achieve good detection ability for abnormal cervical cells,which can help pathologists to conduct rapid screening of focal areas in clinical practice.It has a negative and positive classification ability for WSI,which can be used for hunt management in the early screening of diseases.
作者
杨香山
张欣欣
蔡东兴
车志龙
陈辰
YANG Xiangshan;ZHANG Xinxin;CAI Dongxing;CHE Zhilong;CHEN Chen(Department of Pathology,Shandong Provincial Third Hospital,Jinan 250031;Shandong Flag Information Technology Co.,Ltd.,Jinan 250100)
出处
《北京生物医学工程》
2025年第4期347-356,共10页
Beijing Biomedical Engineering
基金
山东大学临床医学科技创新计划(202019102)资助。
