基于卷积神经网络的PET/CT多模态图像识别研究

摘要

The convolutional neural network is used to recognize CT-PET-PET/CT three modality medical images,it's convenient for hospital to storage and administrate images uniformly,and help doctors retrieve rapidly.Firstly,the feasibility of CNN for the PET/CT multimodal image recognition is discussed.Secondly,the relation between model parameters(iterations,batchsize) and recognition rate,training time is analyzed.Thirdly,the network layers,feature maps and the kernel how to influence the network training and classification effects is discussed.Finally,the experiments show that CNN is feasible to the PET/CT multimodal image recognition,input image size and information complexity should be taken into account while constructing the optimal CNN model for the specific problem,besides,appropriate parameters are choosen to reduce the time complexity with the high recognition rate.%将卷积神经网络用于CT、PET、PET/CT三种模态的医学影像分类识别,为医院统一存储管理影像数据和医护人员快速检索提供便利.首先探讨卷积神经网络对于PET/CT多种模态图像识别的可行性,其次探讨模型参数(迭代次数、批量大小)对网络识别率和训练时间的影响,然后改变CNN模型结构,探讨网络层数、特征图数量和卷积核大小对网络训练和分类效果的影响.实验表明:卷积神经网络对于PET/CT多模态图像识别取得了良好的效果,针对特定问题需要综合图像大小和信息的复杂程度构建最优的CNN模型,在保证高识别率的同时,可以选择合适的参数降低时间复杂度.