一种轻量级脑胶质瘤分割模型

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1671-6841（2025）05-0054-08

DOI：10.13705/j.issn.1671-6841.2023250

Abstract：In recent years，Transformer based automatic segmentation models for brain gliomas greatly improved in performance，but there were stillsome problems such as a large number of parameters，high computational power requirements， large sample requirements，and training diffculties，which could affect the actual deployment of the models. Therefore，a lightweight MRI glioma segmentation model MNATSPNet was proposed. Firstly，a lightweight component MobileNAT was designed to reduce the complexity of Transformer's multi-head self-attention through the adjacency attention mechanism. Secondly， the L1 structured pruning operation was introduced to remove the redundant parameters of the multi-head adjacency attention and fedforward neural network layer in MobileNAT. The experimental results demonstrated that MobileNAT and structured pruning operations could effectively reduce parameters of the model while maintaining stable segmentation performance. Finally，compared with other classc models， MNATSPNet achieved the best results.

Key words： brain glioma segmentation； attention mechanism；Transformer； lightweight； structuredpruning

0 引言

脑肿瘤可生长于脑血管、神经等颅内组织，严重威胁患者的生命健康。（剩余11953字）