摘要
目的 针对传统卷积神经网络(CNN)在MRI图像中空间位置信息建模不足及通道特征感知能力薄弱的问题,提升脑肿瘤识别的准确率与鲁棒性。方法 提出融合坐标卷积(CoordConv)与多级残差注意力机制的改进型ResNet50模型(CoResAttNet),通过在初始层引入CoordConv增强空间感知能力,并在各残差阶段嵌入注意力模块以强化关键通道特征的表达。结果 在两个脑肿瘤MRI公开数据集上进行四分类实验,其在准确率、精确率、召回率和F1分数上分别达到99.72%、99.70%、99.69%、99.69%及98.26%、98.24%、98.25%、98.24%。结论 CoResAttNet模型在脑肿瘤分类任务中具备较好泛化能力与识别性能,能为智能辅助诊断提供有效支持。
Objective To address the limitations of conventional convolutional neural networks in modeling spatial positional information and channel feature perception in MRI images,aiming to improve the accuracy and robustness of brain tumor recognition.Methods An improved ResNet50 model(CoResAttNet)integrating coordinate convolution(CoordConv)and multi-level residual attention mechanisms is proposed.The model enhances spatial awareness by introducing CoordConv in the initial layer and strengthens the representation of critical channel features by embedding attention modules in each residual stage.Results Four-class classification experiments conducted on two publicly available brain tumor MRI datasets demonstrate that CoResAttNet achieves accuracy,precision,recall,and F1-score of 99.72%,99.70%,99.69%,and 99.69%,respectively,on one dataset,and 98.26%,98.24%,98.25%,and 98.24%on the other.Conclusion CoResAttNet model exhibits good recognition performance and generalization ability in brain tumor classification tasks,providing effective support for intelligent computer-aided diagnosis.
作者
闭应洲
甘秋静
霍雷刚
刘善锐
BI Yingzhou;GAN Qiujing;HUO Leigang(Guangxi Key Laboratory of Human-Computer Interaction and Intelligent Decision Making,Nanning Normal University,Nanning,Guangxi Zhuang Autonomous Region 530199,P.R.China)
出处
《临床放射学杂志》
北大核心
2026年第2期213-220,共8页
Journal of Clinical Radiology
基金
国家自然科学基金项目(编号:62067007)
广西学位与研究生教改课题项目(编号:JGY2023236)。
