深度学习在抗核抗体荧光核型识别中的应用初探

Application of deep learning in immunofluorescence images recognition of antinuclear antibodies

目的开发抗核抗体荧光核型的人工智能识别系统雏形,以期满足临床实验室工作中对抗核抗体(ANA)图像免疫荧光模式自动判读的日常需求。方法回顾性分析上海交通大学医学院附属新华医院2020年4月1日至2021年12月31日进行ANA项目检测结果为阳性的荧光图像,3位资深的技术人员独立、平行对荧光图像进行结果判读确定ANA核型结果,并按照ANA荧光模式国际共识荧光核型分类标准进行标注。对常见核型分7个标签组:粗颗粒型、细颗粒型、均质型、核仁型、着丝点型、核点型及核膜型;每个标签组数据按9︰1的比例采用随机数的方法随机分为训练集、验证集,在深度学习框架pytorch1.7上,以ResNet-34图像分类网络为基础构建卷积神经网络训练平台,建立ANA核型自动判读系统。模型建立后单独设立测试集,以模型预测概率高低排序输出判断结果,并对测试集数据进行判读,以人工判读结果作为金标准。应用准确率、精确度、查全率、F1指数等参数评估模型的性能。结果经分割标注后共获取23138张建模图像。共训练了7个模型,比较了不同的算法、图像处理方式和增强方法对模型的影响,选出准确率最高的ResNet-34模型作为最佳模型,其在测试集分类准确率达到93.31%,精确率为91.00%,查全率为90.50%,F1指数为91.50%。测试集的判读结果显示模型对7种核型的ANA图像的识别与人工判读的总体符合率达90.05%,其中对核仁型的判读精确度最高,符合率均达到了100.00%。结论本研究构建的ANA自动判读系统雏形已基本具备判读ANA荧光核型的能力,对常见的、典型的、单一荧光核型的具有一定的准确性。

Objective To develop a prototype artificial intelligence immunofluorescence image recognition system for classification of antinuclear antibodies in order to meet the growing clinical requirements for an automatic readout and classification of immunof luorescence patterns for antinuclear antibody(ANA)images.Methods Immunofluorescence images with positive results of ANA in Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine from April 2020 to December 2021 were collected.Three senior technicians independently and in parallel interpreted the Immunofluorescence images to determine the ANA results.Then the images were labeled according to the ANA International Consensus on Fluorescence Patterns(ICAP)classification criteria.There were 7 labeled groups:Fine speckled,Coarse speckled,Homogeneous,nucleolar,Centromere,Nuclear dots and Nuclear envelope.Each group was randomly divided into training dataset and validation dataset at a ratio of 9∶1 by using random number table.On the deep learning framework PyTORCH 1.7,the convolutional neural network(CNN)training platform was constructed based on ResNet-34 image classification network,and the automatic ANA recognition system was established.After the model was established,the test set was set up separately,the judgment results of the model were output by ranking the prediction probability,with the results of the 2 senior technicians was taken as"golden standard".Parameters such as accuracy,precision,recall and F1-score were used as indicators to evaluate the performance of the model.Results A total of 23138 immunofluorescence images were obtained after segmentation and annotation.A total of 7 models were trained,and the effects of different algorithms,image processing and enhancement methods on the model were compared.The ResNet-34 model with the highest accuracy andswas selected as the final model,with the classification accuracy of 93.31%,precision rate of 91%,and recall rate of 90%and F1-score of 91%in the test set.The overall coincidence rate between the model and manual interpretation was 90.05%,and the accuracy of recognition of nucleolus was the highest,with the coincidence rate reaching 100%in the test set.Conclusion The current AI system developed based on deep learning of the ANA immunofluorescence images in the present study showed the ability to recognize ANA pattern,especially in the common,typical,simple pattern.