【摘要】 目的基于深度学习算法开发一种能够分析肺组织病理切片并实时给出辅助诊断的人工智能(AI)模型。方法回顾性收集2019年7月至2020年2月同济大学附属同济医院胸外科和呼吸科及上海市肺科医院超声科通过手术或超声/CT引导下穿刺活检获得的952例患者肺部病变组织H-E染色病理切片,包括鳞状细胞癌254例(26.68%)、腺癌278例(29.20%)、其他恶性肿瘤47例(4.94%)、良性病变373例(39.18%)。依据病理亚型将切片分层随机采样并按照5∶3∶2的比例分入训练集(476例)、验证集(286例)和测试集(190例)。训练集的每张切片随机截取10张放大400倍的单视野图像用于YOLO v3和Google Inception v3网络的训练,分别开发良恶性病变区域分割模块和病理亚型分类模块,最终构成双模块并联的AI模型。采用同样的方式在验证集中截取单视野图像,用于AI模型与病理医师的诊断能力对比。在测试集中,2名病理医师在显微镜下阅读整张切片进行诊断,其中1名会额外获得AI模型的辅助诊断信息,比较两者的诊断能力。结果验证集共有2 860张单视野图像,其中恶性病变图像1 700张(59.44%),良性病变图像1 160张(40.56%);AI模型鉴别良恶性病变的灵敏度优于病理医师[100%(1 700/1 700)vs 99.65%(1 694/1 700),χ~2=4.167,P=0.031],病理亚型分类准确度与病理医师相当[95.52%(2 732/2 860)vs 94.30%(2 697/2 860),P>0.05],但前者分割恶性病变区域与金标准的重叠率、鉴别良恶性病变的特异度和准确度均低于后者[(92.72±12.75)% vs (95.42±6.99)%,t=7.628,P=0.001;97.67%(1 133/1 160)vs 99.31%(1 152/1 160),χ~2=12.000,P=0.001;99.06%(2 833/2 860)vs 99.51%(2 846/2 860),χ~2=4.364,P=0.037]。测试集共包含190张病理切片,其中恶性病变117张,良性病变73张;获得AI模型辅助的病理医师和未获得辅助的病理医师在良恶性鉴别准确度[100%(190/190) vs 99.47%(189/190),P>0.05]和病理亚型分类准确度[96.84%(184/190) vs 93.68%(178/190),P>0.05]方面差异均无统计学意义;但前者的诊断用时短于后者[(12.53±10.93)s vs(79.95±40.02)s,t=28.939,P<0.01]。结论基于深度学习算法的AI模型能够协助病理医师快速分析H-E染色的肺组织病理切片,在不降低准确度的前提下有效提高了灵敏度和工作效率。更多还原

【Abstract】 Objective To develop an artificial intelligence (AI) model for analyzing pathological sections of lung tissue and providing real-time auxiliary diagnosis based on the deep learning algorithm.Methods Pathological sections of lung lesion tissues from 952 patients obtained by surgery or ultrasound/computed tomography guided biopsy in the Department of Thoracic Surgery and Respiration,Tongji Hospital and the Department of Ultrasound,Shanghai Pulmonary Hospital,Tongji University from Jul.2019 to Feb.2020 were retrospectively collected,including 254 (26.68%) cases of squamous cell carcinoma,278 (29.20%) cases of adenocarcinoma,47 (4.94%) cases of other malignant tumors,and 373 (39.18%) cases of benign lesions.According to the pathological subtypes,the slices were divided into training set (476 cases),validation set (286 cases) and test set (190 cases) using stratified random sampling (5∶3∶2).In each slice of the training set,10 single field images (×400) were randomly captured for the training of YOLO (you only look once) v3 and Google Inception v3 networks,and then the module for segmenting benign and malignant regions and the module for classifying pathological subtypes were developed.Finally,the AI model with two parallel modules was constructed.In the validation set,single field images were captured in the same way to compare the diagnostic ability between the model and the pathologists.In the test set,two pathologists read the whole slices under the microscope for diagnosis,and one of them would get additional auxiliary diagnostic information from the AI model to compare their diagnostic ability.Results There were 2 860 single field images in the validation set,including 1 700 (59.44%) of malignant lesions and 1 160 (40.56%) of benign lesions.The sensitivity of benign and malignant differentiation of the model was better than that of pathologists (100% [1 700/1 700]vs 99.65% [1 694/1 700],χ~2=4.167,P=0.031) and the accuracy of pathological subtype classification was similar to that of pathologists (95.52% [2 732/2 860]vs 94.30% [2 697/2 860],P>0.05).However,the overlap rate of segmented area with the gold standard ([92.72±12.75]% vs [95.42±6.99]%,t=7.628,P=0.001),specificity (97.67% [1 133/1 160]vs 99.31% [1 152/1 160],χ~2=12.000,P=0.001) and accuracy (99.06% [2 833/2 860]vs 99.51% [2 846/2 860],χ~2=4.364,P=0.037) of benign and malignant differentiation of the model were lower than those of pathologists.The test set consisted of 190 pathological sections,including 117 malignant lesions and 73 benign lesions.There were no significant differences between the pathologists who assisted by AI model and who did not in the accuracy rates of benign and malignant differentiation (100% [190/190]vs 99.47% [189/190],P>0.05) or classification of pathological subtypes (96.84% [184/190]vs 93.68% [178/190],P>0.05).However,it took significantly less time for the pathologist to diagnose with the AI model ([12.53±10.93]s vs [79.95±40.02]s,t=28.939,P<0.01).Conclusion The AI model based on deep learning algorithm can help pathologists quickly analyze lung tissue pathological sections stained by H-E staining,which can effectively improve the sensitivity and work efficiency without reducing the accuracy.更多还原