【作者】 张超；

【导师】 丁熠；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 对于医学图像分割任务来说,传统的手工方法依托于医生的经验知识,不仅耗时耗力而且分割精度也没有保证。而随着计算机技术的发展,依托于深度学习的自动化分割方法在各个领域显示出了自己独特的优势,于是有了将深度学习与医学图像相结合来实现自动化医学图像分割的技术。在此背景之下,本文以深度学习为工具探索生成式对抗网络在脑肿瘤分割中的应用,主要工作如下:(1)从多尺度特征的角度出发实现了一个基于并行多尺度的生成对抗网络模型,该网络是一种带有注意力机制并且包含多个残差块的多尺度并行结构,能够缓解在网络深度不断加深的过程中出现的梯度弥散与网络失效问题,允许网络同时提取输入数据不同侧重点的特征信息,通过多尺度推理融合来利用这些特征信息,从而帮助网络模型提升对细节信息的敏感度,提高对于脑肿瘤不同区域的分割精度。(2)在研究如何提升生成式对抗网络对小目标分割精度的过程中实现了细粒度提取模块(FEM),该模块在一定程度上减少了网络编码过程中不必要的语义信息丢失,能够深层次地提取到更多的细粒度信息用于加强局部信息与全局信息的几何约束,这些特点使得在深度神经网络中可以有更多精细化特征来描述目标区域,进而提高网络模型在相关任务中的分割表现。(3)从分阶段处理的角度出发实现了一种新颖的二阶段生成式对抗网络To Sta GAN,旨在解决在网络模型中信息丢失的问题,并将其应用于脑肿瘤图像分割领域。该网络模型在分割的过程中构建了一个从粗到细的过程,整个过程分为两个阶段,第一阶段的网络接收输入数据并输出粗分割结果,第二阶段的网络在细粒度提取模块(FEM)所提取的深层语义信息的帮助下,进一步对粗分割结果进行优化,在生成器与鉴别器互相对抗的过程中完成从粗到细的分割过程。更多还原

【Abstract】 For medical image segmentation tasks,the traditional manual method relies on the doctor’s experience and knowledge,which is not only labor intensive,but also the segmentation accuracy is not guaranteed.With the development of computer technology,the automatic segmentation method based on deep learning has shown its own unique advantages in various fields,so there is a technology that combines deep learning and medical images to realize automatic medical image segmentation.In this context,this thesis adopts deep learning as a tool to explore the application of generative adversarial networks in brain tumor segmentation.The main work is as follows:(1)A parallel multi-scale-based generative adversarial network model is implemented from the perspective of multi-scale features.This network is a multi-scale parallel structure with an attention mechanism and multiple residual blocks,which can alleviate the problems of gradient dispersion and network failure that appear in the process of continuous deepening of the network depth.It allows the network to extract the feature map with different focus from the input data at the same time.And multi-scale inference fusion helps the network model to improve the sensitivity of detailed information and the segmentation accuracy of different regions of brain tumors.(2)The fine-grained extraction module(FEM)is implemented in the process of studying how to improve the accuracy of the generative adversarial network for segmentation of small targets.This module can reduce unnecessary loss of semantic information in the coding process to a certain extent.And more fine-grained information can be extracted in-depth to strengthen the geometric constraints of local and global information.These characteristics make it possible to have more refined features in the deep neural network to describe the target area and improve the performance of the network model in related tasks.(3)A novel two-stage generative adversarial network To Sta GAN is implemented from the perspective of staged processing,which aims to solve the problem of information loss in the network model and is applied to the field of brain tumor segmentation.The model constructs a process from coarse-to-fine in the process of segmentation,which firstly to generate coarse segmentation results in the one stage network.And in the second stage,the coarse results are optimized to gradually achieve a fine result,which is under the guidance of features obtained from the proposed fine-grained extraction module.更多还原