基于深度学习的医学脑图像分割算法研究

【作者】 刘艳；

【导师】 张相芬；

【作者基本信息】 上海师范大学 ， 信息与通信工程， 2022， 硕士

【摘要】 医学图像分割能够为医疗诊断和人类医学研究提供可靠的依据,所以对医学图像分割进行更深入的研究对手术设计、疾病诊断以及愈后评估等临床实践具有重要意义。医学图像分割任务就是按照实际需求将医学图像划分成不同的区域,并将其中感兴趣的部分区域精确地标记出来,而这些感兴趣部分提取的精确与否决定了提供的辅助诊断依据是否可靠。因此准确的分割出感兴趣区域十分重要。本文主要研究在深度学习框架下用于医学图像分割的算法,尤其考虑了医学脑MRI图像中对比度低,分割精度差等问题,提出了两种基于深度神经网络的医学图像分割算法,并通过在公认的数据集上进行实验,证明了所提出网络具有优异的分割性能。本文在医学图像分割课题上的研究,主要有以下两点贡献:(1)针对现有的主流3D-UNet分割模型分割精度差,训练效率较低等问题,提出一种基于3D-UNet和交叉重构注意力机制的MRADE-Net网络分割模型。为了能够有效的关注到图像中对分割更有利的特征信息,该模型提出了交叉重构注意力模块,它可以在融合多模态数据信息的同时,也能对特征进行筛选从而突出重要信息,增强网络对图像特征表达的性能;另外,为了减少下采样过程中信息的丢失,设计了一种特征差分模块,它可以将下采样中丢失的信息传输到上采样中并且不造成信息冗余;除此之外,网络中还引入了多级深度监督机制,加速网络的训练。实验结果说明了MRADE-Net在参数量方面基本与3D-Unet持平的情况下,具有更好的Dice系数,分割精度更高。(2)单一模态图像无法描述脑图像的完整特征,从而造成对特征表达能力低下的问题,另外在网络的解码器中高层特征存在信息表达不准确的现象,针对以上问题,提出了一种多模态特征重构融合倒金字塔网络MCRAIP-Net。该网络模型采用特征金字塔的思想,将不同层级、不同尺寸的特征进行融合,并基于其融合特征实现脑组织的分割,充分利用上下文信息提取脑图像细节特征。此外,还设计了一种多模态交叉重构编码器来对同一层级的不同模态的特征进行融合,从而提高分割精度及网络性能。值得一提的是,在该重构编码器中,为了使多模态特征更充分地融合,提出了双通道交叉重构注意力模块。实验结果表明,本文提出的MCRAIP-Net模型在网络训练效率、分割准确率以及网络结构相似性上都有更优的效果。更多还原

【Abstract】 Medical image segmentation can provide a reliable basis for medical diagnosis and human medical research,so further research on medical image segmentation is of great significance to clinical practice such as surgical design,disease diagnosis and prognosis evaluation.The task of medical image segmentation is to divide the medical image into different regions according to actual needs,and accurately mark the interesting part of the region,and the accuracy of the extraction of these interesting parts determines whether the provided auxiliary diagnosis basis is reliable.Therefore,it is very important to accurately segment the region of interest.This paper mainly studies the algorithms for medical image segmentation under the framework of deep learning,especially considering the problems of low contrast and poor segmentation accuracy in medical brain MRI images,and proposes two medical image segmentation algorithms based on deep neural networks.Experiments on well-established datasets demonstrate the excellent segmentation performance of the proposed network.The research on medical image segmentation in this paper mainly has the following two contributions:(1)In view of the problems of poor segmentation accuracy and low training efficiency of the existing mainstream 3D-UNet segmentation models,a MRADE-Net network segmentation model based on 3D-UNet and cross-reconstruction attention mechanism is proposed.In order to effectively pay attention to the feature information that is more favorable for segmentation in the image,the model proposes a crossreconstruction attention module,which can not only fuse multi-modal data information,but also filter the features to highlight important information.Enhance the performance of the network for image feature expression;in addition,in order to reduce the loss of information during the down-sampling process,a feature difference module is designed,which can transfer the information lost in the down-sampling to the up-sampling without causing information redundancy;In addition,a multi-level deep supervision mechanism is also introduced into the network to speed up the training of the network.The experimental results show that MRADE-Net has better Dice coefficient and higher segmentation accuracy when the parameter quantity is basically the same as that of 3DUnet.(2)A single modal image cannot describe the complete features of the brain image,resulting in the problem of low feature expression ability.In addition,the high-level features in the decoder of the network have inaccurate information representation.Aiming at the above problems,a multi-modality feature reconstruction fusion inverted pyramid network MCRAIP-Net is proposed.The network model adopts the idea of feature pyramid,fuses feature of different levels and sizes,and realizes the segmentation of brain tissue based on the fusion features,and makes full use of context information to extract detailed features of brain images.In addition,a multi-modal cross-reconstruction encoder is designed to fuse the features of different modalities at the same level,thereby improving the segmentation accuracy and network performance.It is worth mentioning that,in this reconstruction encoder,a dual-channel crossreconstruction attention module is proposed to make the multimodal features fused more fully.The experimental results show that the MCRAIP-Net model proposed in this paper has better effects on network training efficiency,segmentation accuracy and network structure similarity.更多还原