摘要
由于传统卷积神经网络CNN(Convolutional Neural Networks)受卷积核尺度的限制,容易丢失磁共振成像MRI(Magnetic Resonance Imaging)脑肿瘤图像的全局信息,而且卷积、池化的过程会导致网络浅层的部分信息丢失,造成基于CNN的脑肿瘤分割特征信息不足,分割精度不高。针对上述问题,提出一种具有全局通路,同时结合网络浅层信息的多通路CNN模型,用来完成多模态MRI脑部神经胶质瘤的全自动分割任务。算法主要思想:将三维多模态MRI图像沿轴向切片化,在相同序列的切片上按比例选取尺度为33×33像素的图像块,得到训练集;将训练集图像块输入多通路CNN模型进行训练;将测试集输入训练好的模型,将脑肿瘤从脑部MRI图像中正确分割出来,并具体划分为坏死、水肿、增强和非增强四种区域,利用模型评估参数Dice系数、敏感度(Sensitivity)系数和特异度(Specificity)系数评测模型的质量。实验结果表明,该方法操作简单,能够有效地完成脑肿瘤的分割任务。
Convolutional Neural Networks(CNN)tends to lose the global information of magnetic resonance imaging(MRI)brain tumor images due to the limitation of convolution kernel scale.Convolution and pooling process can lead to the loss of some information in the shallow layer of the network,resulting in insufficient segmentation information of brain tumor based on CNN and poor segmentation accuracy.To solve the above problems,a multi-channel CNN model with global access and shallow information in the network was proposed to accomplish the automatic segmentation of multi-modality MRI brain gliomas.The algorithm sliced the 3D multi-modality MRI image into axial slices,and selected the scale of 33×33 image blocks on the same slice sequence to get the training set.The training set image block was input into the multi-channel CNN model for training.The test set was input into the trained model,and the brain tumor was correctly segmented from the MRI images of the brain,and was divided into four areas of necrosis,edema,enhancement and non-enhancement.The model was used to assess the quality of the model by evaluating the parameters Dice coefficient,sensitivity coefficient and specificity coefficient.Experimental results show that the proposed method is simple and effectively accomplished the task of brain tumor segmentation.
作者
朱婷
王瑜
肖洪兵
曹利红
Zhu Ting;Wang Yu;Xiao Hongbing;Cao Lihong(Key Laboratory of Food Safety Big Data Technology,School of Computer and Information Engineering, Beijing Technology and Business University,Beijing 100048,China)
出处
《计算机应用与软件》
北大核心
2018年第4期220-226,共7页
Computer Applications and Software
基金
国家自然科学基金项目(61671028)
北京市自然科学基金面上项目(4162018)
北京市委组织部"高创计划"青年拔尖人才培养项目(2014000026833ZK14)
北京市属高等学校高层次人才引进与培养计划项目(CIT&TCD201504010)
2017年研究生科研能力提升计划项目
