卷积神经网络与Vision Transformer在胶质瘤中的研究进展OA

Gliomas pose significant challenges to traditional diagnosis and treatment due to their high heterogeneity,strong invasiveness,and poor prognosis.The introduction of deep learning(DL)technology has opened up a new avenue for their precise diagnosis and treatment,among which convolutional neural network(CNN)and Vision Transformer(ViT)are core tools.CNN inherently excels in local feature extraction(e.g.,tumor edges,texture details)through hierarchical convolution operations,while ViT stands out in global context modeling(e.g.,cross-regional heterogeneity of tumors,multimodal correlations)based on the self-attention mechanism.The fusion strategy of CNN and ViT integrates local fine-grained features with global associated information,demonstrating remarkable advantages in addressing clinical dilemmas such as blurred glioma boundaries and cross-modal data heterogeneity.This article reviews the research progress of CNN and ViT in key clinical tasks of gliomas,including detection and segmentation,pathological grading,molecular subtyping,and prognosis assessment.It elaborates on their principles,individual applications,and fusion strategies.Furthermore,it discusses the prevailing challenges in the field,such as the heavy reliance on annotated data and insufficient model interpretability,and outlines promising future research directions,including the development of lightweight architectures,the advancement of self-supervised learning paradigms,and the promotion of multi-omics integration.This review thereby provides a systematic reference for the intelligent diagnosis of gliomas.