基于注意力机制网络的多实例学习框架实现慢性胃炎多项病理指标的自动识别  

A novel attention fusion network-based multiple instance learning framework to automate diagnosis of chronic gastritis with multiple indicators

作　　者：黄丹 王奕 游庆华 王鑫 张敬谊[4] 丁偕[4] 张伯强 崔浩阳 赵嘉旭 盛伟琪 Huang Dan;Wang Yi;You Qinghua;Wang Xin;Zhang Jingyi;Ding Xie;Zhang Boqiang;Cui Haoyang;Zhao Jiaxu;Sheng Weiqi(Department of Pathology,Fudan University Shanghai Cancer Center Department of Oncology,Shanghai Medical College,Fudan University Institute of Pathology,Fudan University,Shanghai 200032,China;Shanghai Engineering Research Center of Artificial Intelligence Technology for Neoplastic Diseases,Shanghai 200032,China;Department of Pathology,Shanghai Pudong Hospital,Fudan University Pudong Medical Center,Shanghai 201399,China;Wonders Information Co.Ltd,Shanghai 201112,China;Shanghai Foremost Medical Technology Co.Ltd,Shanghai 201112,China)

机构地区：[1]复旦大学附属肿瘤医院病理科,复旦大学上海医学院肿瘤学系,复旦大学病理研究所,200032 [2]上海肿瘤疾病人工智能工程技术研究中心,200032 [3]上海市浦东医院暨复旦大学附属浦东医院病理科,201399 [4]万达信息股份有限公司,上海201112 [5]上海爱可及医疗科技有限公司,201112

出　　处：《中华病理学杂志》2021年第10期1116-1121,共6页Chinese Journal of Pathology

基　　金：上海市经济和信息化委员会2019年上海市人工智能创新发展专项支持项目(2019-RGZN-01017)。

摘　　要：目的应用注意力机制网络的多实例学习(Attention-MIL)框架技术,实现慢性胃炎多项指标的自动识别。方法收集2018年1月1日至12月31日复旦大学附属肿瘤医院诊断为胃炎活检病例1015例和上海市浦东医院诊断为胃炎活检病例115例,所有病理切片经扫描仪进行数字化处理,转化为全载玻片成像(whole slide imaging,WSI),WSI标签依据胃炎病理报告,包含活动性、萎缩和肠化3项指标。所有的WSI分为训练集、单一测试集、混合测试集和外部测试集,Attention-MIL模型在3个测试集上评价自动识别的准确性。结果Attention-MIL模型在240例WSI单一测试集上的受试者工作特征曲线下面积(AUC)值分别为:“活动性”0.98,“萎缩”0.89,“肠化”0.98,3项指标的平均准确率为94.2%。模型在117例WSI混合测试集上的AUC值分别为:“活动性”0.95,“萎缩”0.86,“肠化”0.94,3项指标的平均准确率为88.3%。模型在115例WSI外部测试集上的AUC值分别为:“活动性”0.93,“萎缩”0.84,“肠化”0.90,3项指标的平均准确率为85.5%。结论在慢性胃炎的人工智能辅助病理诊断中,Attention-MIL模型的诊断准确性非常接近病理医师的诊断结果,这种弱监督下的深度学习模式适于病理人工智能技术的实际应用。Objective To explore the performance of the attention-multiple instance learning(MIL)framework,an attention fusion network-based MIL,in the automated diagnosis of chronic gastritis with multiple indicators.Methods A total of 1015 biopsy cases of gastritis diagnosed in Fudan University Cancer Hospital,Shanghai,China and 115 biopsy cases of gastritis diagnosed in Shanghai Pudong Hospital,Shanghai,China were collected from January 1st to December 31st in 2018.All pathological sections were digitally converted into whole slide imaging(WSI).The WSI label was based on the corresponding pathological report,including"activity""atrophy"and"intestinal metaplasia".The WSI were divided into a training set,a single test set,a mixed test set and an independent test set.The accuracy of automated diagnosis for the Attention-MIL model was validated in three test sets.Results The area under receive-operator curve(AUC)values of Attention-MIL model in single test sets of 240 WSI were:activity 0.98,atrophy 0.89,and intestinal metaplasia 0.98;the average accuracy of the three indicators was 94.2%.The AUC values in mixed test sets of 117 WSI were:activity 0.95,atrophy 0.86,and intestinal metaplasia 0.94;the average accuracy of the three indicators was 88.3%.The AUC values in independent test sets of 115 WSI were:activity 0.93,atrophy 0.84,and intestinal metaplasia 0.90;the average accuracy of the three indicators was 85.5%.Conclusions To assist in pathological diagnosis of chronic gastritis,the diagnostic accuracy of Attention-MIL model is very close to that of pathologists.Thus,it is suitable for practical application of artificial intelligence technology.

关 键 词：胃炎 人工智能 模式识别 自动 信号处理 计算机辅助 注意力 

分 类 号：R573[医药卫生—消化系统] TP391.41[医药卫生—内科学]