ADHD Classification With Low-Frequency Fluctuation Feature Map Based on 3D CBAMe
ADHD Classification With Low-Frequency Fluctuation Feature Map Based on 3D CBAMe
复制标题
DOI：
10.1145/3563737.3563749
复制
发表时间：
2022-08
期刊：
影响因子：
--
通讯作者：
Lihua Su;Sei-ichiro Kamata
中科院分区：
文献类型：
--
作者：
Lihua Su;Sei-ichiro Kamata
注意缺陷多动障碍（ADHD）是一种常见的青少年神经发育障碍。一些优秀的ADHD自动诊断系统从磁共振图像中提取特征。研究人员已经表明，功能磁共振成像数据提供了多动症大脑活动的具体测量方法。本文提出了一种用于ADHD诊断的低频波动特征图生成方法，该方法可以突出fMRI特征的判别部分。然而，提取的特征映射仍然存在冗余信息。因此，我们增加了关注机制，使其更加关注局部信息。为了成功地将注意机制应用于卷积神经网络（CNN），并将其与三维fMRI特征图匹配，我们将卷积块注意模块（CBAM）从二维平面扩展到三维几何空间。在此基础上，设计了基于三维CBAM的单模态三维CNN，通过低频波动特征图对ADHD进行诊断。我们的模型在ADHD-200数据集上进行了评估，获得了75.83%的最新分类准确率。同时，我们的模型还简化了多模态方法的特征提取模块和分类模块。
Attention deficit hyperactivity disorder (ADHD) is a common neurodevelopmental disorder in teenagers. Some excellent ADHD automatic diagnosis system extracted features from magnetic resonance image (MRI). Researchers have shown fMRI data offers specific measure of ADHD brain activity. In this paper, we propose a low-frequency fluctuation feature map generation approach for ADHD diagnosis, which can highlight the discriminative parts of fMRI features. However, the extracted feature maps still have redundant information. So we add the attention mechanism which can pay more attention to the local information. In order to successfully apply the attention mechanism to convolutional neural network (CNN) and match it to 3D fMRI feature maps, we extend convolutional block attention module (CBAM) from 2D plane to 3D geometric space. After that, we design a single modality 3D CNN based on 3D CBAM to diagnosis ADHD via low-frequency fluctuation feature map. Our model is evaluated on ADHD-200 dataset and it obtains the state-of-the-art classification accuracy of 75.83%. At the same time, our model also simplifies the feature extraction module and the classification module of multi-modality method.