摘要
脑肿瘤精准分割至关重要,但传统卷积神经网络因局部感受野限制难以建模磁共振成像(magnetic resonance imaging,MRI)中的长距离依赖,影响对异质性高、边界模糊肿瘤的分割精度。Transformer凭借全局自注意力机制为此提供了新思路。本文综述了Transformer在脑肿瘤MRI分割中的进展,重点分析了Transformer模型在层次化注意力、编解码结构、残差连接等关键技术上的改进,探讨了多模态融合、模态缺失应对、轻量化设计及注意力机制本身的创新策略;尽管Transformer显著提升了精度,仍面临数据稀缺、模态缺失鲁棒性、类别不平衡、计算成本高和可解释性不足等挑战,未来需聚焦数据高效利用、模态弹性建模、拓扑感知优化、轻量化与可解释性增强等方向。本文系统梳理了Transfomer在脑肿瘤MRI图像分割领域的研究现状,总结了目前研究的局限性并指出未来的研究方向,本文旨在为深入理解其技术演进、核心挑战与发展方向提供系统性参考。
Accurate segmentation of brain tumors is crucial,but traditional convolutional neural networks are difficult to model long-range dependencies in magnetic resonance imaging(MRI)due to local receptive field limitations,which affects the segmentation accuracy of tumors with high heterogeneity and blurred boundaries.Transformer provides a new approach for this through its global self-attention mechanism.This article reviews the progress of Transformer in brain tumor MRI segmentation,focusing on analyzing the improvements of Transformer models in key technologies such as hierarchical attention,encoder-decoder structures,and residual connections,and exploring innovative strategies for multimodal fusion,handling missing modalities,lightweight design,and the attention mechanism itself;although Transformers have significantly improved accuracy,they still face challenges such as data scarcity,robustness to modal loss,class imbalance,high computational costs,and insufficient interpretability,necessitating future focus on efficient data utilization,modal elasticity modeling,topology-aware optimization,lightweight and interpretability enhancement,and other directions.This article systematically reviews the current research status of Transformer in the field of brain tumor MRI image segmentation,summarizes the limitations of current research,and points out the future research directions.The aim is to provide a systematic reference for a deeper understanding of its technological evolution,core challenges,and development trends.
作者
陈雷
李光宇
杨锋
蔡婧欣
高梦谣
CHEN Lei;LI Guangyu;YANG Feng;CAI Jingxin;GAO Mengyao(Department of Assets and Equipment,the Affiliated Hospital of Shandong University of Chinese Medicine,Jinan 250014,China;School of Medical Information Engineering,Shandong University of Chinese Medicine,Jinan 250355,China)
出处
《磁共振成像》
北大核心
2025年第8期181-187,200,共8页
Chinese Journal of Magnetic Resonance Imaging
