In order to solve the problem of difficulty in capturing global features of three-dimensional convolutional neural network (3D CNN) in the classification and prediction of neuroimaging data, a multi-layer perceptron mixer (MLP-Mixer) that effectively models global relationships was introduced, and a feature fusion based hybrid model was proposed aimed at improving the classification performance of neuroimaging data by combining the advantages of 3D CNN and MLP-Mixer. Firstly, in the feature extraction stage, 3D CNN was used to extract local spatial features from resting-state functional magnetic resonance imaging (rs-fMRI) data, while MLP-Mixer was used to capture global relational features. Secondly, in the fusion classification module, the scaled dot-product self-attention technique was used to combine spatial features with global relationship features, dynamically balancing and integrating the importance of different features, enhancing the expressive power of the model, and completing the classification task through a fully connected layer. The experimental results show that the hybrid model has achieved a maximum improvement of 4.12% and 5.16% in classification accuracy for binary and multi-class classification tasks of neural imaging data, respectively. It is also significantly better than other network models in terms of recall rate, F1 score, and other indicators. This indicates that the hybrid model proposed can effectively improve classification accuracy and provide a new technological approach for automated classification of neuroimaging data, with broad potential for engineering applications.
在深度学习与神经影像学结合的研究中,已有研究利用3D CNN或多层感知混合器(MLP-Mixer)模型成功地对神经影像数据进行分类,展示了处理神经元激活移位和缩放的能力。例如,Thanh等[1]基于3D CNN计算机辅助诊断方法,通过从静息态fMRI(rs-fMRI)数据中提取独立成分分析(Independent component analysis,ICA)特征,阿尔茨海默病(AD)与正常对照组(NC)分类准确率达到85.27%;Hanh等[4]利用3D CNN模型提取任务相关信息,成功地对AD的实时fMRI进行五类分类,分类准确率可达78.5%;Li等[5]提出端到端的数据驱动模型(C3d-LSTM),结合3D CNN,从fMRI图像序列中提取每一体积的三维静态图像空间特征,实现了97.37%的AD和NC分类准确率;He等[6]基于自监督学习(Self-supervised learning,SSL)提出了通道注意MLP-Mixer网络(SCAMLP-Net),用于脑电图(Electroencephalogram,EEG)分类,通过有效地学习时间和空间信息,显示出优异的分类性能;Qiu等[7]结合完全卷积网络(Fully convolutional network,FCN)和传统MLP,分析结构性MRI(sMRI)数据并生成高分辨率的疾病概率图,达到了神经科医生级别的AD状态诊断精度;Zhou等[8]提出的ME-Mixer CNN模型,通过整合ME-Mixer模块到主干CNN中,对COVID-19和ISIC-2019数据集进行分类,表现优于其他方法;Zhang等[9]基于MLP-Mixer架构建立了3D-Mixer模型,通过对sMRI和PET数据集进行AD与NC的分类,在曲线下面积(Area under the curve,AUC)、精确率-召回率曲线下面积(Area under the precision-recall curve,AUPR)和准确率(Accuracy,ACC)指标表现均出色。
本文共纳入了吉林大学第一医院放射科登记的年龄在60至85岁之间的老年人,总计有69名受试者的静息态功能性磁共振数据(rs-fMRI)被采用。这一年龄段的选择是基于其高发的神经退行性疾病风险,特别是60岁以上人群中,AD及其早期阶段的发生率显著增加。所有参与者在数据采集前提供了书面知情同意书,在部分AD患者无法自身提供知情同意的情况下,近亲在参与前给予同意。为了保证数据的准确性,参与者的选择标准为:①无脑损伤史;②无神经精神疾患;③无磁共振检查禁忌物(包括心脏起搏器植入患者、体内有移植物或金属异物患者);④右利手。排除标准为:①有严重基础疾病患者;②幽闭症患者;③无法配合实验者[12]。另外,参与者也完成了简易精神状态量表评分(Mini-mental state examination,MMSE)和临床痴呆评定量表(Clinical dementia rating,CDR)的神经心理学测验评估,并记录得分。最终,经过MMSE和CDR测试选定的受试者信息为NC有18名、AD有17名、MCI有18名、SCD有16名,具体信息如表1所示。其中,NC(正常对照组)为基线对照,受试者无明显的脑萎缩或异常病变;AD(阿尔茨海默病组),患者大脑海马体明显萎缩;MCI(轻度认知障碍组),患者海马体萎缩程度介于NC与AD组之间,尚未达到AD的严重程度;SCD(主观认知下降组),患者脑结构无明显萎缩或异常,但主观报告有认知下降的体验。
缩放点积自注意力机制是一种改进的点积注意力方法,旨在减轻输入向量对注意力权重的影响,并融合不同特征。本文利用该机制实现特征融合和分类。核心步骤是:首先通过查询向量 Q 与键向量 K 之间的点积,衡量二者相似性,然后通过Softmax函数将点积结果转换为概率分布,从而确定值向量 V 的加权重要性,最后,经过加权求和生成输出[15]。其公式如下:
式中: Q 、 K 、 V 分别为查询矩阵、键矩阵和值矩阵;为键向量的维度,作为缩放因子引入,为了防止高维度情况下的点积值过大,避免出现梯度消失或梯度爆炸的现象。
如图5所示,缩放点积自注意力的执行过程首先是接收前面阶段提取到的特征。具体而言,3D CNN负责捕捉局部空间特征,而MLP-Mixer则用于提取全局特征。两个模块的输出在经过线性变换(Linear transformation)后,分别映射为 Q 、 K 和 V。其中,线性变换通过全连接层(FC)实现,旨在将不同维度的数据统一到注意力机制的计算空间中。
然后,查询向量 Q 与键向量 K 之间的点积结果通过缩放因子进行缩放,之后经过Softmax处理后,点积结果被转换为归一化的概率分布,用以计算值向量 V 的权重。通过对 V 进行加权求和,生成上下文向量,该向量结合了 Q 、 K 和 V 三者所包含的全局特征信息。为充分利用3D CNN与MLP-Mixer提取到的特征,生成的上下文向量会与3D CNN的原始输出进行拼接,从而实现全局和局部特征的结合。拼接后的特征通过线性层进行变换,压缩维度,最终经过全连接层进行处理,完成分类任务并输出最终的分类结果。
ThanhN D, SeungjunR, IqbalN M Q, et al. 3D-deep learning based automatic diagnosis of Alzheimer's disease with joint MMSE Prediction using resting-state fMRI[J]. Neuroinformatics, 2020, 18(1): 71-86.
[2]
WangM L, LianC F, YaoD R, et al. Spatial-temporal dependency modeling and network hub detection for functional MRI analysis via convolutional-recurrent network[J]. IEEE Transactions on Biomedical Engineering, 2019, 67(8): 2241-2252.
[3]
WarrenS L, MoustafaA A. Functional magnetic resonance imaging, deep learning, and Alzheimer's disease: a systematic review[J]. Journal of Neuroimaging: Official Journal of the American Society of Neuroimaging, 2022, 33(1): 5-18.
[4]
HanhV, Hyun-ChulK, MinyoungJ, et al. fMRI volume classification using a 3D convolutional neural network robust to shifted and scaled neuronal activations[J]. NeuroImage, 2020, 223: 117328-117328.
[5]
LiW, LinX, ChenX. Detecting Alzheimer's disease based on 4D fMRI: an exploration under deep learning framework[J]. Neurocomputing, 2020, 388: 280-287.
[6]
HeY, LuZ, WangJ, et al. A self-supervised learning based channel attention MLP-mixer network for motor imagery decoding[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2022, 30: 2406-2417.
[7]
QiuS, JoshiP S, MillerM I, et al. Development and validation of an interpretable deep learning framework for Alzheimer's disease classification[J]. Brain: a Journal of Neurology, 2020, 143(6): 1920-1933.
[8]
ZhouZ, IslamM T, XingL. Multibranch CNN with MLP-mixer-based feature exploration for high-performance disease diagnosis[J]. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(6): 7351-7362.
[9]
ZhangZi-chao, ZhaoXing-zhong, DongGui-ying, et al. Improving Alzheimer's disease diagnosis with multi-modal PET embedding features by a 3D multi-task MLP-mixer neural network[J]. IEEE Journal of Biomedical and Health Informatics, 2023, 27(8): 4040-4051.
WangWei, SunYu-jie, WangXin. Lightweight Frequency and spatial feature fused multi-scale remote sensing scene classification network[J]. Journal of Jilin University (Engineering and Technology Edition), 2025, 55(10): 3361-3371.
[12]
LiuY B, YueL, XiaoS F, et al. Assessing clinical progression from subjective cognitive decline to mild cognitive impairment with incomplete multi-modal neuroimages[J]. Medical Image Analysis, 2022, 75: 102266-102266.
[13]
ChenZ P, MaX, ChenR, et al. A reparametrized CNN model to distinguish Alzheimer's disease applying multiple morphological metrics and deep semantic features from structural MRI[J]. Frontiers in Aging Neuroscience, 2022, 14: 856391-856391.
WangXue-zhi, LiQing-liang, LiWen-hui. Spatio-temporal model of soil moisture prediction integrated with transfer learning[J]. Journal of Jilin University(Engineering and Technology Edition), 2022, 52(3): 675-683.
JiangSheng, WangYi-Di, XieRui-Lin, et al. Traffic accident anticipation baed on spatial-temporal relational learning and convolutional gated recurrent network[J]. Journal of Jilin University (Engineering and Technology Edition), 2025, 55(3): 954-962.
[18]
SohnG, ZhangN, OlukotunK. Implementing and optimizing the scaled dot-product attention on streaming dataflow[J/OL]. [2024-11-21].
LiYun-hong, WangMei, SuXue-ping, et al. Road extraction from remote sensing images combining attention and context fusion[J]. Journal of Jilin University (Engineering and Technology Edition), 2025, 55(12): 4034-4044.
[21]
ZhouP, XieX, LinZ, et al. Towards understanding convergence and generalization of AdamW[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(9): 6486-6493.
[22]
SimaG, AhmadS. Predicting conversion from MCI to AD by integration of rs-fMRI and clinical information using 3D-convolutional neural network[J]. International Journal of Computer Assisted Radiology and Surgery, 2022, 17(7): 1245-1255.
[23]
MohammedA, TianfuW, AhmedE. Fusing multimodal and anatomical volumes of interest features using convolutional auto-encoder and convolutional neural networks for Alzheimer's disease diagnosis[J]. Frontiers in Aging Neuroscience, 2022, 14: 812870-812870.