基于有效感受野和注意力融合机制的脑肿瘤全自动分割  

Fully-automatic brain tumor segmentation based on effective receptive field and attention fusion mechanism

作　　者：邹祥 王瑜[1] 肖洪兵[1] 杨迪 ZOU Xiang;WANG Yu;XIAO Hongbing;YANG Di(School of Computer and Artificial Intelligence,Beijing Technology and Business University,Beijing 100048,China;China CAMC Engineering Co.,Ltd.,Beijing 100048,China)

机构地区：[1]北京工商大学计算机与人工智能学院,北京100048 [2]中工国际工程股份有限公司,北京100048

出　　处：《中国医学物理学杂志》2024年第5期563-570,共8页Chinese Journal of Medical Physics

基　　金：北京市自然科学基金-北京市教育委员会科技计划重点项目(KZ202110011015)。

摘　　要：深度学习在医学图像分割领域取得了显著成果,但其在脑肿瘤分割任务中,仍面临感受野不足、冗余信息过多、信息丢失等问题;为此,本研究提出一种基于编-解码结构的脑肿瘤分割网络模型(EAU-Net)。EAU-Net采用有效感受野拓展模块和注意力融合模块改善脑肿瘤分割网络感受野不足与冗余信息过多带来的不利影响;同时,引入基于倒残差结构的瓶颈重采样模块,有效避免上下采样时造成的信息损失,并采用深度卷积降低网络的计算量。在BraTS2020数据集上的实验结果表明,EAU-Net获得最优的分割精度,验证了其在脑肿瘤分割任务中的可行性和有效性。Despite significant achievements of deep learning in medical image segmentation,there are challenges for brain tumor segmentation using deep learning,such as insufficient receptive field,excessive redundant information,and information loss.To address these issues,a novel brain tumor segmentation network model(EAU-Net)is proposed based on encoder-decoder structure.EAU-Net incorporates an effective receptive field expansion block and an attention fusion module to minimize the adverse effects caused by insufficient receptive field and excessive redundant information which often occurred in the current brain tumor segmentation network.Additionally,a bottleneck resampling module based on inverted residual structure is introduced to effectively avoid information loss during upsampling and downsampling,while deep convolutions are used to reduce computational complexity.Experimental results on the BraTS2020 dataset reveal that EAU-Net achieves the highest segmentation accuracy,demonstrating its feasibility and effectiveness for brain tumor segmentation.

关 键 词：脑肿瘤分割 EAU-Net 有效感受野拓展模块 注意力融合模块 倒残差结构 

分 类 号：R318[医药卫生—生物医学工程] TP181[医药卫生—基础医学]