基于Swin Transformer的脑肿瘤图像自动分类方法  

Automatic Classification Method Based on Swin Transformer for Brain Tumor Images

作　　者：焦德琼 冯磊 刘伟志 宁春玉[1] JIAO Deqiong;FENG Lei;LIU Weizhi;NING Chunyu(School of Life Science and Technology,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学生命科学技术学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2024年第4期122-129,共8页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技发展计划项目(20220101123JC)。

摘　　要：针对脑肿瘤图像分类准确率低的问题,提出一种基于Swin Transformer的改进模型,命名为ClassSwin。该模型在初始阶段嵌入卷积Stem,以提高对图像局部信息的获取能力;利用ClassSwin Block提升模型捕获全局信息的能力。在ClassSwin Block中,引入SCAB,使用一对基于空间注意力和通道注意力的相互依赖分支,来有效地学习空间和通道的脑肿瘤特征信息。ClassSwin与其他分类模型在Kaggle数据集上进行了脑肿瘤四分类任务的实验验证,其准确率、精确率、召回率、特异性和F1分数分别达到99.24%、99.28%、99.18%、99.74%和99.23%。实验结果证明了该方法有助于医疗专家准确诊断脑肿瘤类型,为未来脑肿瘤分类领域的研究提供了基准。Aiming at the problem of low accuracy in brain tumor image classification,an improved model based on Swin Transformer is proposed,named Classification Swin Transformer(ClassSwin).The model embeds convolutional Stem in the initial stage to improve the acquisition of local information of the image,and utilize ClassSwin Block to enhance the model's ability to capture global information.In ClassSwin Block,SCAB is introduced which uses a pair of interdependent branches based on spatial attention and channel attention to effectively learn brain tumor feature information of space and channel.ClassSwin and other classification models were experimentally validated on the Kaggle dataset for brain tumor four classification tasks,with accuracy,precision,recall,specificity,and F1scores of 99.24%,99.28%,99.18%,99.74% and99.23%,respectively.The experimental results have proven that this method is helpful for medical experts to accurately diagnose brain tumor types,providing a benchmark for future research in the field of brain tumor classification。

关 键 词：脑肿瘤分类 深度学习 TRANSFORMER Swin Transformer 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]