基于病理大切片的深度卷积神经网络对于直肠癌脉管癌栓的预测价值  

PREDICTIVE VALUE OF DEEP CONVOLUTIONAL NEURAL NETWORK BASED ON LARGE HISTOLOGICAL SECTIONS FOR VASCULAR TUMOR EMBOLI IN RECTAL CANCER

作　　者：鞠宜衡 张宪祥[1] 郑龙波 赵鹏[2] 卢云[1,3,4] JU Yiheng;ZHANG Xianxiang;ZHENG Longbo;ZHAO Peng;LU Yun(Department of Gastrointestinal Surgery, The Affiliated Hospital of Qingdao University, Qingdao 266003, China)

机构地区：[1]青岛大学附属医院胃肠外二科,山东青岛266003 [2]青岛大学附属医院病理科 [3]山东省数字医学与计算机辅助手术重点实验室 [4]青岛大学附属医院海南分院

出　　处：《精准医学杂志》2022年第2期131-136,共6页Journal of Precision Medicine

基　　金：国家自然基金青年基金项目(81802473)。

摘　　要：目的基于病理大切片数字图像构建预测直肠癌脉管癌栓的深度卷积神经网络(DCNN)模型,以提高临床医生对该病的预测能力。方法收集2019年1月—2019年10月青岛大学附属医院西海岸院区120例患者(内部数据集)、市南院区+崂山院区40例患者(外部数据集)的直肠癌术后病理标本,经石蜡包埋、切片、染色、扫描后共获得2400张病理大切片HE染色数字图像。将内部数据集患者按7∶3比例随机分为训练组(84例)和测试组(36例),用以构建和测试预测脉管癌栓的DCNN模型,以外部数据集患者的数字图像验证该模型。通过准确率、精准率、灵敏度、特异度、受试者工作特征(ROC)曲线及曲线下面积(AUC)、精度-召回率曲线(PR)及曲线下面积(AP)等指标来测试和验证模型的性能。结果本研究成功建立了脉管癌栓预测的DCNN模型。该模型的测试结果显示,准确率为85.2%,精准率为83.8%,灵敏度为84.6%,特异度为84.8%,AUC值为0.83,AP值为0.80;验证结果显示,准确率为84.7%,精准率为83.1%,灵敏度为84.3%,特异度为84.1%,AUC值为0.82,AP值为0.78,其中在验证结果中,模型的自动预测时间、准确率、灵敏度及特异度与病理医师人工HE染色阅片结果相比,均有显著差异(χ^(2)=5.028~6.842,t=-73.84,P<0.05)。结论基于病理大切片数字图像构建的预测直肠癌脉管癌栓的DCNN模型与人工HE染色阅片相比对直肠癌患者脉管癌栓具有较高的预测能力,对临床医生诊断直肠癌脉管癌栓具有一定的辅助作用。Objective To construct a deep convolutional neural network(DCNN)model based on digital images of large histological sections,and to improve the ability of clinicians to predict the disease.Methods We collected the postoperative pathological specimens of rectal cancer from 120 patients(internal dataset)from the West Coast Hospital and 40 patients(external dataset)from the Shinan Hospital and Laoshan Hospital of The Affiliated Hospital of Qingdao University from January to October in 2019.A total of 2400 digital images of HE-stained large sections were obtained after paraffin embedding,sectioning,staining,and scanning.The cases in the internal dataset were randomly divided into training group(84 cases)and test group(36 cases)in a ratio of 7∶3 to construct and test a DCNN model for predicting vascular tumor emboli.The model was validated with the external dataset.Finally,the performance of the model was evaluated with accuracy,precision,sensitivity,specificity,a receiver operating characteristic(ROC)curve,area under the ROC curve(AUC),a precision-recall curve(PR),and area under the PR curve(AP).Results A DCNN model for vascular tumor embolus prediction was basically established.Testing results showed that this model had an accuracy of 85.2%,a precision of 83.8%,a sensitivity of 84.6%,a specificity of 84.8%,an AUC value of 0.83,and an AP value of 0.80.Validation results showed that the model had an accuracy of 84.7%,a precision of 83.1%,a sensitivity of 84.3%,a specificity of 84.1%,an AUC value of 0.82,and an AP value of 0.78;the model showed significant differences in accuracy,sensitivity,specificity,and time to diagnosis compared with reading by the pathologist(χ^(2)=5.028-6.842,t=-73.84,P<0.05).Conclusion The DCNN model based on digital images of large sections has a higher ability to predict vascular tumor emboli compared with reading by pathologists,and can be helpful for clinicians to diagnose vascular tumor emboli in rectal cancer.

关 键 词：直肠癌 脉管癌栓 病理大切片 人工智能 深度卷积神经网络 神经网络 计算机 模型构建 大数据 诊断 

分 类 号：R735.3[医药卫生—肿瘤]