【摘要】 目的 开发基于DenseNet算法对膜性肾病(MN)肾小球钉突病理图像进行分类的人工智能模型，研究人工智能模型是否能辅助病理科医师在MN病理检测中发现钉突这一细微结构，提高病理科医师对MN的病理诊断水平。方法 选取2014-2019年山西医科大学附属第二医院收治的MN患者肾组织针刺活检病理切片1 250张，经筛选符合要求的六胺银(PASM)染色病理切片1 150张，选择诊断为MN病理分期Ⅱ期的PASM染色病理图像127张。前期实验通过Cascade R-CNN网络识别并检测肾小球。由高年资病理科医师对切割后的肾小球进行分类，将含有钉突样改变的肾小球视为钉突阳性，共492张图像；将不含有钉突样改变的肾小球视为钉突阴性，共523张图像。使用基于深度学习的DenseNet分类网络构建对肾小球钉突进行分类的人工智能模型。将数据集以8∶2分为训练集和测试集，模型性能通过测试数据集进行评估。使用经训练的DenseNet模型对图像进行测试。通过灵敏度、特异度和受试者工作特征曲线的曲线下面积(AUC)对训练后的模型进行评估。结果 基于DenseNet模型是正确检测到的是否有钉突的肾小球并对其进行二分类。根据测试结果得到召回率为98.00%,精确度为92.45%,准确率为95.00%,F1为95.15%。DenseNet模型表现出高性能，AUC为0.97。结论 DenseNet对肾小球钉突的二分类获得了较高的召回率、准确率和灵敏度，但精确度和特异度尚需进一步提高才能更好地辅助病理科医师诊断MN。更多还原

【Abstract】 Objective To develop an artificial intelligence model based on DenseNet algorithm to classify the pathological images of glomerular spikes in membranous nephropathy(MN),and to study whether the artificial intelligence model can assist the pathologists to find the fine structure of spikes in the pathological detection of MN,so as to improve the diagnosis level of pathological physicians on MN.Methods A total of 1 250 pathological sections of renal tissue needle biopsy from the patients with MN in the Second Affiliated Hospital of Shanxi Medical University during 2014-2019 were selected, 1 150 periodic acid-silver metheramine(PASM) staining pathological sections after screening met the requirement and 127 PASM staining pathological sections diagnosed as the stage II MN were selected.The earlier stage experiment identified and detected the glomerulus by the Cascade R-CNN network.The cut glomeruli were classified by a senior pathologist, and the glomerulus containing the spike-like change was considered as spikes positive, with a total of 492 images.The glomerulus without the spike-like change were considered as spikes negative, with a total of 523 images.The DenseNet classification network based on deep learning was used to construct an artificial intelligence model to classify the glomerular spikes.The data set was divided into the training dataset and testing dataset with a ratio of 8∶2,and the model performance was evaluated through the testing the dataset.The trained DenseNet model was used to test the image.The post-training model was evaluated by the sensitivity, specificity and area under the curve(AUC)of the receiver operating characteristic(ROC) curve.Results Based on the DenseNet model, the glomeruli with and without spikes were correctly detected and dichotomized.According to the test results, the obtained recall rate was 98.00%,the precision was 92.45%,the accuracy was 95.00% and F1-score was 95.15%.The DenseNet model showed the high performance with AUC of 0.97.Conclusion The dichotomization of glomerular spikes by DenseNet has achieved the high recall rate, accuracy and sensitivity, but the precision and specificity need to be further improved in order to better assist the pathological doctors in the diagnosis of MN.更多还原