【作者】 张斌；

【导师】 刘伯强；

【作者基本信息】 山东大学 ， 生物医学工程（专业学位）， 2021， 硕士

【摘要】 胶质瘤是最常见的原发性脑肿瘤,有着高病发率、高死亡率的特点,严重危害着人类的生命健康。随着医学成像技术的发展,医学影像已成为辅助医生进行医学诊断和研究的重要手段,其中核磁共振成像(Magnetic resonance imaging,MRI)技术,由于其具有非入侵性、良好的空间分辨率和软组织分辨率等优点,被广泛应用于脑成像。脑肿瘤分割有助于医生对患者做出早期诊断、治疗规划和预后评估,但是手动分割费时费力,并且会受到医生技术水平的影响。因此,开发精确的脑肿瘤自动分割技术具有重要的意义。然而,脑肿瘤具有高异质性,在多模态MRI脑影像中表现为灰度值的不均匀性和形状的不规则性,并且脑肿瘤分割存在严重的类不平衡问题,因此开发精确可靠的自动分割算法是一项具有挑战性的工作。近年来,许多基于深度学习的方法被应用于脑肿瘤分割,取得了不错的效果。针对深度学习在脑肿瘤分割中的应用,本文深入探讨了相关技术,并提出了基于多重注意力机制的脑肿瘤分割模型。具体的研究内容和创新工作如下:1、提出了一种新型的基于多重注意力(Multiple Attention U-Net,MAU-Net)的脑肿瘤分割网络,将注意力充分运用于空间信息、通道信息和尺度信息。注意力的使用不仅提取了更丰富的语义信息,而且更加关注小区域脑肿瘤的信息,从而提高了脑肿瘤的分割效果。2、在3D U-Net的最后一层编码器中,添加了双注意力模块(Dual attention module,DAM),其中并联了一个空间注意力块(Spatial attention block,SAB)和一个通道注意力块(Channelattentionblock,CAB),以突出显示显著的特征信息,同时消除无关和嘈杂的特征响应。在上采样的过程中,提出了一个多重注意力门模块(Multi-attention gate module,MAGM),融合了来自编码器的特征图、双注意力模块的输出,以及当前解码器的特征图。通过融合多尺度上下文信息,充分利用多尺度图像特征以实现脑肿瘤的精确分割。3、在训练过程中使用了缩放、对比度增强和高斯噪声等数据增强方式以缓解因数据集偏小而引发的过拟合现象。并使用了一种基于Dice损失和交叉熵损失的融合损失函数。交叉熵损失可以帮助网络加速收敛,保持训练稳定。Dice损失能够缓解类不平衡现象。4、在BraTS2018数据集中进行了广泛的实验。通过消融实验验证了各个模块的有效性,并与其它最先进的方法相比,实验结果表明,MAU-Net取得了较好的分割结果。更多还原

【Abstract】 Glioma is the most common primary brain tumor with high morbidity and mortality,which seriously endangers human life and health.With the development of medical imaging technology,medical imaging has become an important tool to assist doctors in medical diagnosis and researches.As one of these techniques,Magnetic Resonance Imaging(MRI)is widely used for brain imaging for its advantages of non-invasiveness,good spatial resolution and soft tissue resolution.Brain tumor segmentation helps physicians to make early diagnosis,treatment planning and prognosis assessment of patients,but manual segmentation is time-consuming and laborious,and can be affected by the skill level of physicians.Therefore,it is important to develop accurate automatic brain tumor segmentation techniques.However,brain tumors are highly heterogeneous and exhibit heterogeneity of gray values and irregularity of shapes in multimodal MRI brain images,and there is a serious class imbalance problem in brain tumor segmentation,so it is a challenging task to develop accurate and reliable automatic segmentation algorithms.In recent years,many deep learning-based methods have been applied to brain tumor segmentation with good results.For the application of deep learning in brain tumor segmentation,this paper discusses the related techniques in depth and proposes a brain tumor segmentation model based on multiple attention mechanisms.The specific research components and innovative work are as follows:1、A novel brain tumor segmentation network based on Multiple Attention U-Net(MA-Net)is proposed,which fully utilizes attention to spatial information,channel information and scale information.The use of attention not only extracts richer semantic information,but also focuses more on the information of small brain tumors.thus improving the segmentation of brain tumors.2、In the final layer of the 3D U-Net encoder,a Dual attention module(DAM)is added,in which a Spatial attention block(SAB)and a Channel attention block(CAB)are concatenated to highlight the salient feature information while eliminating irrelevant and noisy feature responses.In the upsampling process,a Multi-attention gate module(MAGM)is proposed to fuse the feature maps from the encoder,the output of the dual-attention module,and the feature maps of the current decoder to make full use of the multi-scale image features for accurate segmentation of brain tumors.3、Data enhancements such as scaling,contrast enhancement and Gaussian noise are used in the training process to alleviate the overfitting phenomenon caused by the small dataset.And a fusion loss function based on Dice loss and Cross Entropy loss is used.The Cross Entropy loss can help the network to accelerate the convergence and keep the training stable.Dice loss can alleviate the class imbalance phenomenon.4、Extensive experiments are conducted in the BraTS2018 dataset.The effectiveness of each module is verified by ablation experiments and compared with other state-of-the-art methods,the experimental results show that MAU-Net achieved better segmentation results.更多还原