基于深度学习算法的病理学图片淋巴细胞浸润检测

Deep learning-based lymphocyte infiltration detection on pathological images

背景与目的:深度学习方法可用来在病理切片上开展淋巴细胞的自动分割和检测。本研究探讨使用变分自编码模型预训练的方法对病理学图片进行淋巴细胞浸润检测的性能,以及去除肿瘤坏死区域对模型性能的影响。方法:使用变分自编码模型(variational auto-encoder,VAE)先在来自肿瘤基因组图谱(the Cancer Genome Atlas, TCGA)数据库中的无标注的TCGA-COAD和TCGA-READ病理切片图像上进行预训练,获得一个自编码预训练模型,再在已标注淋巴细胞的公开数据集上进行淋巴细胞浸润检测的模型训练。为避免与坏死的区域相混淆,还训练了分割肿瘤坏死区域的Unet模型,以去除肿瘤坏死区域对淋巴细胞识别的影响。分析经过变分自编码模型预训练的淋巴细胞浸润检测模型在训练集和测试集上的受试者工作特征曲线(receiver operating characteristic curve,ROC)的曲线下面积(area under curve,AUC)及其95%置信区间(confidence interval,CI)。结果:检测模型在训练集上AUC为0.979(95%CI:0.978~0.980),准确率为92.5%(95%CI:92.3%~92.6%),Kappa值为0.849,灵敏度为0.908,特异度为0.941,精确率为0.939,召回率为0.908,F1为0.923。验证集的AUC为0.968(95%CI:0.964~0.972),准确率为91.3%(95%CI:90.6%~92.0%),Kappa值为0.826,灵敏度为0.898,特异度为0.928,精确率为0.925,召回率为0.898,F1为0.912。Resnet18模型在已标注的数据集上直接训练的结果为:验证集的准确率为83.2%(95%CI:82.2%~84.1%),Kappa值为0.664,灵敏度为0.823,特异度为0.840,精确率为0.838,召回率为0.823,F1为0.830。分割肿瘤坏死区域的Unet模型最终在训练集上的DICE值为0.78,在验证集上为0.76。去除坏死区域后本文提出的变分自编码模型预训练的淋巴细胞浸润检测模型的预测性能获得了一定的提升,在验证集上的AUC从0.968(95%CI:0.964~0.972)提升至0.971(95%CI:0.968~0.975)。准确率为92.4%(95%CI:91.7%~93.0%),Kappa值为0.849,灵敏度为0.928,特异度为0.921,精确率为0.921,召回率为0.928,F1为0.925。结论:采用变分自编码模型预训练的方法,对淋巴细胞浸润的病理学图片进行浸润检测,可获得比直接训练更优的模型表现;并且通过Unet去除肿瘤坏死区域的影响能够进一步提高模型的性能。

Background and purpose:Deep learning methods can be used for automatic segmentation and detection of lymphocytes on pathological images.This study aimed to assess the performance of using variational autoencoding pre-training method for lymphocyte infiltration detection on pathological images,as well as the impact of removing tumor necrosis regions on model performance.Methods:Using variational autoencoding(VAE)pre-training method,pre-training was performed on a large number of unlabeled pathological images from the Cancer Genome Atlas(TCGA)database(TCGA-COAD and TCGA-READ)to obtain an auto-encoding pre-training model,and then a classifier model of lymphocyte infiltration was trained on the public data samples.To avoid confusion with necrotic regions,a Unet segmentation model for tumor necrotic regions was trained to remove the influence of tumor necrotic regions on lymphocyte identification.Results:The lymphocyte infiltration detection model pre-trained with the VAE model had an area under curve(AUC)of 0.979(95%CI:0.978-0.980),an accuracy of 92.5%(95%CI:92.3%-92.6%),a kappa value of 0.849,sensitivity of 0.908,specificity of 0.941,precision of 0.939,recall of 0.908,and F1 of 0.923 under the receiver operating characteristic(ROC)curve on the training set.The AUC for the validation set was 0.968(95%CI:0.964-0.972),the accuracy was 91.3%(95%CI:90.6%-92.0%),kappa value was 0.826,sensitivity was 0.898,specificity was 0.928,precision was 0.925,recall was 0.898,and F1 was 0.912.The results of Resnet18 model on the labeled dataset were as follows:accuracy of the validation set was 83.2%(95%CI:82.2%-84.1%),kappa value was 0.664,sensitivity was 0.823,specificity was 0.840,precision was 0.838,recall was 0.823 and F1 was 0.830.The Unet model that segmented the necrotic regions of the tumors had a final DICE of 0.78 for the training set,and 0.76 for the validation.After removing the necrotic region,the predictive performance of the pre-trained lymphocyte infiltration detection model using the VAE proposed in this article was improved to some extent,with the AUC on the validation set increasing from 0.968(95%CI:0.964-0.972)to 0.971(95%CI:0.968-0.975).The accuracy was 92.4%(95%CI:91.7%-93.0%),kappa value was 0.849,sensitivity was 0.928,specificity was 0.921,precision was 0.921,recall was 0.928,and F1 was 0.925.Conclusion:Using the variational autoencoding model pre-training method to classify the pathological pictures of lymphocyte infiltration can obtain better model performance compared with direct training,and removing the influence of tumor necrosis areas through Unet can further improve the performance of the model.