基于深度学习算法的脑肿瘤CT图像特征分割技术改进

崔仲远 黄伟

摘 要： 针对基于数学形态的脑肿瘤CT图像特征分割技术存在准确率低、分割效果不明确的弊端，提出基于深度学习算法的脑肿瘤CT图像特征分割技术。将可视人体数据集CVH?2作为研究对象，对数据集中的图像实施预处理，对图像四个模态实施卷积分别获取不同模态彼此的差异信息，归一化获取脑肿瘤CT图像多模态3D?CNNs特征。对基于SAE深度学习算法的脑肿瘤CT图像特征分割模型实施二级训练，将脑肿瘤CT图像多模态3D?CNNs特征经过处理后获取的S，V通道数据输入模型实施训练，在第二级训练的过程中把第一级SAE训练得到的权重作为二级训练的原始权重，将一级训练中错误分割的组织结构和沟回作为二次训练的数据集，获取脑肿瘤CT图像特征的准确分割结果。实验结果表明，所提方法在脑肿瘤CT图像特征分割准确率和效率方面具有显著优势。

关键词： 深度学习算法； 脑肿瘤CT图像； 特征分割技术； 多模态3D?CNN； SAE结构； 数据集

中图分类号： TN911.73?34； R739.41 文献标识码： A 文章编号： 1004?373X（2018）16?0092?04

Abstract： In allusion to the disadvantages of low accuracy rate and unobvious segmentation effect of the brain tumor CT image feature segmentation technology based on mathematical morphology， a brain tumor CT image feature segmentation technology based on in?depth learning algorithm is proposed. By taking the visual human body data set CVH?2 as the research object， the preprocessing of the images in data set is performed. The convolution is performed for the four modes of images， so as to obtain the differentiated information of different modes respectively， and obtain the multimodal 3D?CNNs features of brain tumor CT images in normalization. The second?level training is conducted for the feature segmentation model of brain tumor CT images based on the SAE in?depth learning algorithm. The S and V channel data obtained after the processing of the multimodal 3D?CNNs features of brain tumor CT images is input into the model for training. During the process of the second?level training， the weight obtained from the first?level SAE training is taken as the original weight of the second?level training， and the organizational structures and sulci wrongly segmented during the first?level training are taken as the data set of the second?level training， so as to obtain the accurate segmentation results of brain tumor CT image features. The experimental results show that the proposed method has a significant advantage in improving the accuracy rate and efficiency of brain tumor CT image feature segmentation.

Keywords： in?depth learning algorithm； brain tumor CT image； feature segmentation technology； multimodal 3D?CNN； SAE structure； data set

0 引 言

随着现代医疗水平的提高，医学成像技术在日常医疗诊断与医学研究中的作用日益显著，因此对医学诊断图像数据的研究至关重要。脑肿瘤作为时下频繁出现且复杂性较强的肿瘤疾病，已成为医学界重点研究的课题。脑肿瘤的确诊通常以脑CT图像的成像数据分析为依据。准确地分析脑CT图像是判断病人病情的关键步骤。但是医生个人医疗知识的积累、经验水平的差异以及视觉疲劳等不确定因素都会影响对图像结果的正确分析。……