一种基于深度学习多分类模型的肺部细胞病理涂片鉴别诊断方法

A diagnosis approach of lung cytological images based on deep learning multi-classification

目的采用快速在线细胞病理学评估方法进行肺癌中晚期患者的病理评估,是目前常用诊断方法,但存在人工诊断准确率低和细胞病理医生人数不足等问题。本文提出一种基于深度学习的细胞病理涂片多分类方法,以求实现六类常见肺部细胞病理涂片的鉴别诊断。方法本文提出了一种基于CBAM注意力机制增强的ResNet-18网络,以及一种由粗到细的多分类框架,并对深度学习分类方法的特征激活图进行了分析。结果本文共收集了313张肺部Diff-quick染色的细胞病理涂片,其中259张用于训练,54张用于测试。本文所提出方法在正常肺组织、小细胞癌、非小细胞癌、鳞癌、腺癌和类癌共计6种细胞的分类鉴别中取得了准确率为70.4%、精确率为81.5%,召回率为78.2%和F1评分为78.9%的结果。在与金标准的相关性对比中,该模型与高年资细胞病理学医生相当,高于低年资细胞病理学医生。结论本文提出了一种基于深度学习多分类模型的肺部细胞病理涂片鉴别诊断方法,该方法可以协助细胞病理学医生进行肺癌患者的细胞病理涂片诊断,并提高快速在线细胞病理学评估的可行性。

Objective Needle aspiration of lung mass using Rapid On-Site cytologic Evaluation is one of the diagnosis methods for patients in the middle to late stages with poor physical condition.However,the cost of a cytopathologist involved in the procedure is unaffordable for most hospitals.Methods In this study,we proposed a multi-classification approach of lung cytological images based on a ResNet-18 network enhanced by convolutional block attention modules,which could differentiate normal tissue,small cell carcinoma,non-small cell lung cancer,squamous carcinoma,adenocarcinoma and carcinoid tumor.Results A total of 313 patches of diff-quick stained samples were collected.The proposed method was trained on 259 patches and tested on 54 patches.The approach achieved an accuracy,precision,recall and F1-score of 70.4%,81.5%,78.2%,and 78.9%respectively in classifying the six categories.The method performed better than junior cytopathologists,close to senior cytopathologists in its correlation with the gold standard.Conclusions The proposed method could assist cytopathologists in their diagnosis of lung cancer and substantially improve the viability of rapid on-site cytologic evaluation.