基于卷积神经网络的颅内病变类型影像的判别  

Determination of intracranial lesions type image based on convolutional neural network

作　　者：胡明哲 杨永立 

机构地区：[1]武汉科技大学国际学院,武汉430081 [2]武汉科技大学信息科学与工程学院,武汉430081

出　　处：《黑龙江大学自然科学学报》2017年第6期748-756,共9页Journal of Natural Science of Heilongjiang University

基　　金：国家自然科学基金资助项目(61375081)

摘　　要：颅内病变的具体类型直接影响医生所选用的医疗方式,目前颅内病变影像的判别主要依靠医生的经验,易造成误诊。提出了一个基于卷积神经网络的精准影像分类法,通过从医院放射科电子计算机断层扫描设备采集五种较常见病变类型和一种正常颅脑CT图像作为分类的对象进行预处理。创建一个包含3个卷积层、3个池化层、1个完全连接层的卷积神经网络,并对网络采取了Dropout技术优化处理。并用所采集的颅内病变样本对神经网络进行训练和测试。通过实验将改进后的CNN算法与模板比较法及SVM等传统算法进行比较发现,分类结果的准确度明显优于传统算法,平均识别准确率可达93.54%。Specific types of intracranial lesions will affect the selection by medical doctors directly. The way to determine the current image of intracranial lesions still relies on the experience of doctors, which can easily cause misdiagnosing. A precise image classification method is presented based on convolution neural network. The five kinds of common lesion types and a normal brain CT image were selected from the hospital radiology computer tomography equipment as the object of the classification for pretreatment. Then, a convolution neural network is set up with three convolution layers, three pool layers, and one fully connected layer. Dropout technique is adopted to optimize the network. The neural network was trained and tested with the acquired intracranial lesions images. The comparisons of the improved CNN algorithm with the template comparison method, SVM and other traditional algorithms show that the accu- racy of the elassification results of CNN neural network is superior to those traditional algorithm, and the average recognition accuracy rate is able to reach 93.54%.

关 键 词：卷积神经网络(CNN) 计算机断层扫描(CT) 颅内病变 

分 类 号：O436[机械工程—光学工程]