基于双路径特征融合的轻量级脑肿瘤分割网络  

Lightweight Brain Tumor Segmentation Network Based on Dual-Path Feature Fusion

作　　者：李锵 阮方号 关欣 Li Qiang;Ruan Fanghao;Guan Xin(School of Microelectronics,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学微电子学院,天津300072

出　　处：《天津大学学报（自然科学与工程技术版）》2024年第11期1177-1186,共10页Journal of Tianjin University：Science and Technology

基　　金：天津市自然科学基金资助项目(22JCZDJC00220);超声医学与工程国家重点实验室基金资助项目(2022KFKT004).

摘　　要：脑肿瘤是世界上最致命的癌症之一,从三维核磁共振图像(MRI)中快速、准确地分割正常脑组织和恶性肿瘤组织,对临床诊断和手术治疗都至关重要.近年来,基于卷积神经网络的分割架构,特别是3D U-Net架构在脑肿瘤分割方面取得了巨大成功.然而,现有的基于3D U-Net的网络存在下采样信息丢失的问题,缺乏自动聚焦小尺度病灶的能力,编码器和解码器之间的全局上下文信息交互不够充分,并且参数量大、计算开销高.针对上述问题,提出一种双路径特征融合的轻量级脑肿瘤分割网络.首先,该网络通过增加一条分支路径,将原始数据中的低级细节信息添加到编码器的各层中,弥补下采样带来的特征信息损失.其次,提出一种多层金字塔制导模块代替传统的跳跃连接,增强解码器不同层次的全局上下文信息获取能力.最后,在输出层引入多视图级联注意力模块,利用肿瘤不同区域的包含关系,使网络从各个视图自动聚焦小尺度肿瘤区域.在BraTS2020数据集上的实验结果表明,采用0.55×10^(6)参数量和41.21×10^(9)浮点运算次数,该网络在增强肿瘤、全肿瘤和肿瘤核心的Dice系数分别可达78.64%、89.51%和83.77%.此外,在BraTS2018数据集上进一步评估该网络的性能,实验结果表明:该网络在保持较小计算量的同时,显著提高了对脑肿瘤各病灶区域的定位能力和分割性能,在临床实践中具有重要意义.Brain tumors rank among the most lethal forms of cancer worldwide.The fast and accurate segmentation of normal brain tissue and malignant tumor tissue from 3D magnetic resonance imaging(MRI)scans is crucial for clinical diagnosis and surgical treatment.In recent years,convolutional neural network-based segmentation architectures,particularly those using 3D U-Net architectures,have shown exceptional success in this area.However,the existing 3D U-Net-based networks often face significant challenges,including loss of information during down-sampling.They also cannot automatically focus on small-scale foci,fail to incorporate sufficient global context information interaction between encoder and decoder,and possess a large parameter size that induces high computation costs.To overcome these limitations,we propose a lightweight brain tumor segmentation network that utilizes dual-path feature fusion.First,our network enriches each encoder layer with low-level details from the original data through an additional branch path.This design compensates for the loss of feature information typically caused by downsampling.Second,we introduce a multilayer pyramid guidance module.This module serves as a sophisticated alternative to traditional skip connections,significantly enhancing the acquisition of global context information at different levels of the decoder.Lastly,our network features a multiview cascading attention module in the output layer.This module automatically focuses on small-scale tumor regions from each view,utilizing the inclusion relationship between different tumor regions. Our extensive testing on the BraTS2020 data set has yielded promising results,withDice coefficients of 78.64%,89.51%,and 83.77% for enhanced tumor region,whole tumor region,and tumorcore,respectively. These achievements were accomplished using only 0.55×10^(6) parameters and involving 41.21×10^(9) floating point operations. Further validation of the BraTS2018 data set confirmed the network’s superior performancein localizing and segment

关 键 词：核磁共振图像 脑肿瘤分割 轻量级 双路径 多视图级联 多层金字塔制导 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]