张雪媛, 许鸿雁, 董跃明, 刘丹凤, 孙鹏蕊, 颜锐, 崔洪亮, 雷红, 任菲. 基于多尺度注意力机制的真菌显微图像分类方法[J]. 协和医学杂志, 2023, 14(1): 139-147. DOI: 10.12290/xhyxzz.2022-0169
引用本文: 张雪媛, 许鸿雁, 董跃明, 刘丹凤, 孙鹏蕊, 颜锐, 崔洪亮, 雷红, 任菲. 基于多尺度注意力机制的真菌显微图像分类方法[J]. 协和医学杂志, 2023, 14(1): 139-147. DOI: 10.12290/xhyxzz.2022-0169
ZHANG Xueyuan, XU Hongyan, DONG Yueming, LIU Danfeng, SUN Pengrui, YAN Rui, CUI Hongliang, LEI Hong, REN Fei. Fungal Microscopic Image Classification Based on Multi-scale Attention Mechanism[J]. Medical Journal of Peking Union Medical College Hospital, 2023, 14(1): 139-147. DOI: 10.12290/xhyxzz.2022-0169
Citation: ZHANG Xueyuan, XU Hongyan, DONG Yueming, LIU Danfeng, SUN Pengrui, YAN Rui, CUI Hongliang, LEI Hong, REN Fei. Fungal Microscopic Image Classification Based on Multi-scale Attention Mechanism[J]. Medical Journal of Peking Union Medical College Hospital, 2023, 14(1): 139-147. DOI: 10.12290/xhyxzz.2022-0169

基于多尺度注意力机制的真菌显微图像分类方法

Fungal Microscopic Image Classification Based on Multi-scale Attention Mechanism

  • 摘要:
      目的  利用深度学习技术,建立临床常见的侵袭性真菌图像辅助分类模型。
      方法  回顾性收集2020年9月—2021年4月解放军总医院第八医学中心曲霉菌属、酵母菌属和新型隐球菌属真菌感染者的显微镜图像,按7∶1.5∶1.5的比例随机分为训练集、验证集和测试集。使用训练集和验证集图像对改进的MobileNetV2网络结构进行训练和参数调试,构建基于多尺度注意力机制的卷积神经网络(convolutional neural network, CNN)真菌图像11分类模型。以机器鉴定结果为金标准,以查准率、召回率和F1值为指标评价该模型对测试集真菌图像的分类效果。将该模型与5种经典CNN模型进行比较,评价指标包括模型参数量、内存占用量、网络每秒处理的图像数量(frames per second, FPS)、准确率及受试者操作特征曲线下面积(area under the curve,AUC)。
      结果  共纳入真菌显微镜图像7666张,分别包括曲霉菌属、酵母菌属和新型隐球菌属图像2781张、4115张、770张。其中训练集5366张、验证集1150张、测试集1150张。改进的MobileNetV2模型对测试集11种真菌图像具有较高的分类性能,查准率为96.36%~100%,召回率为96.53%~100%,F1值为97.01%~100%。该模型的参数量、内存占用量分别为4.22 M、356.89 M,FPS为573,准确率为(99.09±0.18)%,AUC为0.9944±0.0018,综合性能优于5种经典网络模型。
      结论  本研究提出的真菌图像分类模型,在保持低运算成本的情况下,可获得较高的真菌图像识别能力,其整体性能优于常见的经典模型。

     

    Abstract:
      Objective  To establish a fungal image-assisted classification model using deep learning technology.
      Methods  The microscope images of people infected with Aspergillus, Saccharomyces and Cryptococcus neoformans were retrospectively collected from the Eighth Medical Center of PLA General Hospital from September 2020 to April 2021. The images were randomly divided into training set, validation set and test set according to the ratio of 7∶1.5∶1.5. The improved MobileNetV2 network structure was trained using the training set, a convolutional neural network (CNN) fungal image 11 classification model based on multi-scale attention mechanism was constructed and the parameters were debugged based on the validation set. Machine identification results were taken as the gold standard, the performance of the model on 11 fungal image classification tasks was evaluated, and the results were shown by precision, recall and F1 value. In addition, the performance of the proposed model with 5 classic CNN models were compared, and the results were measured in terms of model parameters, memory usage, frames per second (FPS), accuracy, and area under the curve (AUC) of receiver operating characteristic curve.
      Results  A total of 7666 fungal microscope images were collected, including 2781, 4115, and 770 images of Aspergillus, Saccharomyces, and Cryptococcus neoformans, respectively. Among them, there were 5366 training images, 1150 validation images, and 1150 test images. The improved MobileNetV2 model had high performance for the classification of 11 fungal images in the test set. The precision rate was distributed between 96.36% and 100%, the recall rate was distributed between 96.53% and 100%, and the F1 value was distributed between 97.01% and 100%. The parameters, memory usage, FPS, accuracy, and AUC of the improved MobileNetV2 model were 4.22 M, 356.89 M, 573, (99.09±0.18)%, and 0.9944±0.0018, respectively, and the comprehensive performance was better than 5 kinds of classical networks.
      Conclusion  The proposed fungal image classification model based on the improved MobileNetV2 can obtain higher fungal image recognition ability while maintaining low computational cost, with an overall performance better than classical CNN model.

     

/

返回文章
返回