刘畅, 郑宇超, 谢文倩, 李晨, 李晓晗. 基于深度学习的宫颈原位腺癌图像识别方法[J]. 协和医学杂志, 2023, 14(1): 159-167. DOI: 10.12290/xhyxzz.2022-0109
引用本文: 刘畅, 郑宇超, 谢文倩, 李晨, 李晓晗. 基于深度学习的宫颈原位腺癌图像识别方法[J]. 协和医学杂志, 2023, 14(1): 159-167. DOI: 10.12290/xhyxzz.2022-0109
LIU Chang, ZHENG Yuchao, XIE Wenqian, LI Chen, LI Xiaohan. Image Recognition Method of Cervical Adenocarcinoma in Situ Based on Deep Learning[J]. Medical Journal of Peking Union Medical College Hospital, 2023, 14(1): 159-167. DOI: 10.12290/xhyxzz.2022-0109
Citation: LIU Chang, ZHENG Yuchao, XIE Wenqian, LI Chen, LI Xiaohan. Image Recognition Method of Cervical Adenocarcinoma in Situ Based on Deep Learning[J]. Medical Journal of Peking Union Medical College Hospital, 2023, 14(1): 159-167. DOI: 10.12290/xhyxzz.2022-0109

基于深度学习的宫颈原位腺癌图像识别方法

Image Recognition Method of Cervical Adenocarcinoma in Situ Based on Deep Learning

  • 摘要:
      目的  基于深度学习算法构建宫颈原位腺癌(cervical adenocarcinoma in situ, CAIS)病理图像诊断模型。
      方法  回顾性收集2019年1月至2021年12月中国医科大学附属盛京医院病理科保存的CAIS患者病变组织、慢性宫颈炎患者正常宫颈管腺体病理切片。图像采集后,均按4∶3∶3的比例随机分为训练集、验证集和测试集。使用训练集、验证集数据对VGG16、VGG19、Inception V3、Xception、ResNet50和DenseNet201共6种网络模型进行迁移学习训练和参数调试,构建可识别CAIS病理图像的卷积神经网络二分类模型,并将模型进行组合,构建集成学习模型。基于测试集数据,采用运算时间、准确率、精确率、召回率、F1值、受试者操作特征曲线下面积(area under the curve,AUC)对模型识别CAIS病理图像的性能进行评价。
      结果  共入选符合纳入和排除标准的CAIS患者病理切片104张、慢性宫颈炎患者正常宫颈管腺体病理切片90张。共收集CAIS、正常宫颈管腺体病理图像各500张,其中训练集、验证集、测试集图像分别400张、300张、300张。6种模型中,ResNet50模型的准确率(87.33%)、精确率(90.00%)、F1值(86.90%)及AUC(0.87)均最高,召回率(84.00%)居第2位,运算时间较短(2062.04 s),整体性能最佳,VGG19模型次之,Inception V3与Xception模型的性能最差。6种集成学习模型中,ResNet50与DenseNet201集成模型的整体性能最优,其准确率、精确率、召回率、F1值、AUC分别为89.67%、84.67%、94.07%、89.12%、0.90,VGG19与ResNet50集成模型次之。
      结论  通过深度学习算法构建CAIS病理图像识别模型具有可行性,其中ResNet50模型的整体性能较高。集成学习可提高单一模型对病理图像的识别效果。

     

    Abstract:
      Objective  To construct a pathological image diagnostic model of cervical adenocarcinoma in situ(CAIS) based on deep learning algorithm.
      Methods  Pathological tissue sections of CAIS and normal cervical canal and gland sections of chronic cervicitis stored in the Pathology Department of Shengjing Hospital, China Medical University from January 2019 to December 2021 were retrospectively collected. After image collection, they were randomly divided into training set, validation set and test set with a ratio of 4∶3∶3. The data of training set and validation set were used to conduct transfer learning training and parameter debugging for 6 network models, including VGG16, VGG19, Inception V3, Xception, ResNet50 and DenseNet201, and the convolutional neural network binary classification model that could recognize pathological images of CAIS was constructed. The models were combined to build the ensemble learning model. Based on the test set data, the performance of pathological image recognition of single model and ensemble learning model was evaluated. The results were expressed by operation time, accuracy, precision, recall, F1 score and area under the curve(AUC) of receiver operating characteristic.
      Results  A total of 104 pathological sections of CAIS and 90 pathological sections of normal cervical duct and gland with chronic cervicitis were selected. A total of 500 pathological images of CAIS and normal cervical duct glands were collected, including 400 images of training set, 300 images of validation set and 300 images of test set, respectively. Among the 6 models, ResNet50 model, with the highest accuracy(87.33%), precision(90.00%), F1 score(86.90%) and AUC(0.87), second highest recall(84.00%) and shorter operation time(2062.04 s), demonstrated the best overall performance; VGG19 model was the second; and Inception V3 and Xception model had the worst performance.Among the 6 kinds of ensemble learning models, ResNet50 and DenseNet201 showed the best overall performance, and their accuracy, precision, recall, F1 score and AUC were 89.67%, 84.67%, 94.07%, 89.12% and 0.90, respectively. VGG19 and ResNet50 ensemble model followed.
      Conclusions  It is feasible to construct CAIS pathological image recognition models by deep learning algorithm, among which ResNet50 models has higher overall performance. Ensemble learning can improve the recognition effect on pathological images by single model.

     

/

返回文章
返回