探讨基于卷积神经网络（CNN）的数字病理学模型在肝癌复发预警中的预测价值。

本研究包含4个肝癌队列，队列1和队列3为训练集，队列2和队列4为验证集。队列1、2、3分别有202、179、738例患者，来源于2012年1月至2017年1月浙江大学医学院附属第一医院1 119例肝癌患者。队列4是来自美国的癌症基因组图谱（TCGA）数据库361例患者。首先，在队列1运用7种CNNs（AlexNet、Squeezenet、InceptionV3、GoogleNet、DenseNet201、VGG19和ResNet18）训练肝癌病理切片病灶自动识别模型，区分肿瘤区域与正常组织；选择最佳的神经网络模型，通过迁移学习构建肿瘤区域成分抽提模型，解析肿瘤区域5种不同成分（肿瘤细胞、淋巴细胞、间质、坏死及背景），并在队列2上验证；在队列3上探索这5种成分与肝癌术后复发的相关性，并联合临床特征构建肝癌术后复发预警模型，在队列4进行独立的外部验证。生存分析采用Kaplan-Meier法和Log-rank检验，生存预后影响因素分析采用LASSO-Cox比例风险回归模型。

在队列1中，AlexNet、DenseNet201、VGG19和ResNet18鉴别肝癌病灶的准确率均高于95.0%。在队列2中，神经网络VGG19的准确率最佳，达96.4%，选择VGG19迁移至肿瘤区域成分抽提模型，其分割5种成分的准确率达99.0%。在队列3中，多因素Cox分析显示，高淋巴细胞神经网络评分（LYM）和低间质神经网络评分（STR）与患者术后复发明显相关（HR=0.70，1.38；P<0.05）；基于LYM、STR、AFP、美国癌症联合委员会（AJCC）分期构建术后复发风险的预警模型，在训练集和验证集中，模型可对患者术后复发风险精准分层（χ²=45.06，15.49；P<0.05）。在独立队列4中，肝癌复发预警模型进一步验证，LYM、STR病理特征综合评分可提高AJCC分期的无复发生存预测能力。

基于CNN的肝癌病灶自动识别模型和肿瘤区域成分抽提模型可智能化分割肝癌数字病理切片，实现肝癌术后复发预警，LYM、STR为术后复发相关的新型病理预测因子。

To evaluate the predictive value of digital pathological model based on convolutional neural network (CNN) in the early warning of recurrence of hepatocellular carcinoma (HCC).

This study consisted of 4 HCC cohorts. Cohort 1 and 3 were taken as the training sets, and cohort 2 and 4 as the validation sets. 1 119 HCC patients admitted to the First Affiliated Hospital of Zhejiang University from January 2012 to January 2017 were enrolled and assigned into cohort 1 (n=202), 2 (n=179) and 3 (n=738), respectively. 361 patients from The Cancer Genome Atlas (TCGA) were allocated incohort 4. First, 7 CNNs (AlexNet, Squeezenet, InceptionV3, GoogleNet, DenseNet201, VGG19 and ResNet18) were used to train the automatic recognition model of pathological sections of HCC to distinguish tumor region from normal tissues. The optimal neural network model was chosen. The component extraction model for tumor region was constructed through transfer learning, and 5 different components (tumor cells, lymphocytes, stroma, necrosis and background) in the tumor region were analyzed and were validated by cohort 2. The correlation between these 5 components and postoperative recurrence of HCC was analyzed in cohort 3. A warning model of postoperative recurrence of HCC was constructed based on clinical features, which was subjected to independent external validation in cohort 4. Survival analysis was conducted by Kaplan-Meier method and Log-rank test. Prognostic factors were analyzed by Cox proportional hazards regression model with LASSO method.

In cohort 1, the accuracy rates of AlexNet, DenseNet201, VGG19 and ResNet18 in distinguishing HCC lesions were all above 95.0%. In cohort 2, the accuracy rate of neural network VGG19 was the highest up to 96.4%. When VGG19 was transferred to the component extraction model in the tumor region, the accuracy rate of segmenting 5 components reached 99.0%. Incohort 3, multivariate Cox analysis showed that high lymphocyte (LYM) neural network score and low stroma (STR) neural network score were significantly correlated with postoperative recurrence (HR=0.70, 1.38; P<0.05). Based on LYM, STR, AFP and American Joint Committee on Cancer (AJCC) staging, the early warning model of the risk of postoperative recurrence was constructed. In the training and validation sets, the models could accurately classify the risk of postoperative recurrence risk of patients (χ²=45.06, 15.49; P<0.05). In the independent cohort 4, the early warning model of HCC recurrence was further validated, and the comprehensive scores of LYM and STR pathological features could improve the prediction capability of AJCC staging for recurrence-free survival.

The CNN-based automatic recognition model for HCC lesions and the component extraction model in tumor region can deliver intelligent segmentation of digital pathological sections of HCC and provide early warning of postoperative recurrence of HCC. LYM and STR are novel pathological predictors for postoperative recurrence.