脑状态解码的可解释深度学习新框架
A new interpretable deep learning framework for brain state decoding
卢梅丽 1高资成 1郭兆桦1
作者信息
- 1. 天津职业技术师范大学信息技术工程学院,天津 300222
- 折叠
摘要
针对功能磁共振成像(function magnetic resonance image,fMRI)的脑状态解码未充分挖掘fMRI的时间特征,且需要依赖专家知识进行特征提取的问题,从深度学习可解释性的角度提出一种基于任务态fMRI的智能脑状态解码新框架,先设计基于Inception结构的三维卷积神经网络(Inception-CNN)对任务态fMRI进行分类,再结合类激活映射算法解释分类结果与大脑不同脑区的功能相关性,并提出改进的细粒度类激活算法Guided Score-CAM,采用遮掩校验法验证类激活映射可视化算法的有效性.对所提出的解码框架在4种不同的任务态功能磁共振成像数据中进行测试,结果表明:Inception-CNN分类模型准确度达98%,类激活可视化能够准确映射到分类结果对应的功能脑区,可有效解码大脑任务状态.
Abstract
The brain state decoding of functional magnetic resonance imaging is not able to fully exploit the temporal characteristics of functional magnetic resonance imaging(fMRI),and relies on expert knowledge for feature extraction.This paper proposes a new framework of intelligent brain state decoding based on task-state fMRI from the perspective of deep learning interpretability.It proposes the Guided Score-CAM,an improved fine-grained class activation mapping algorithm,and verifies the effectiveness of the class activation mapping visualization algorithm using the masking test method.The results show that the Inception-CNN classification model achieves an accuracy of 98%.Compared with traditional methods,the brain state decoding framework adopted in this paper can not only accurately realize the model classification,but also accurately locate the activated regions of the brain during corresponding task execution.
关键词
Inception-CNN/任务态功能磁共振图/Guided/Score-CAM/遮掩检验Key words
Inception-CNN/task state functional magnetic resonance imaging/Guided Score-CAM/masking test引用本文复制引用
基金项目
国家自然科学基金(62171312)
天津市教委科研项目(2020KJ114)
出版年
2024
NETL
NSTL
万方数据