Zhang Haifeng, Song Chonghui, Zhao Xiaolong, Wang Fei, Qiu Yunlong, Li Hao, Guo Hongyi
College of Information Science and Engineering, Northeastern University, Shenyang 110819, China.
Division of Psychology, Nanyang Technological University, Singapore S639798, Singapore.
Heliyon. 2024 Jul 10;10(14):e34245. doi: 10.1016/j.heliyon.2024.e34245. eCollection 2024 Jul 30.
Resting-state functional magnetic resonance imaging (rs-fMRI) is a non-invasive neuroimaging technique widely utilized in the research of Autism Spectrum Disorder (ASD), providing preliminary insights into the potential biological mechanisms underlying ASD. Deep learning techniques have demonstrated significant potential in the analysis of rs-fMRI. However, accurately distinguishing between healthy control group and ASD has been a longstanding challenge. In this regard, this work proposes a model featuring a dual-path cross-attention framework for spatial and temporal patterns, named STDCformer, aiming to enhance the accuracy of ASD identification. STDCformer can preserve both temporal-specific patterns and spatial-specific patterns while explicitly interacting spatiotemporal information in depth. The embedding layer of the STDCformer embeds temporal and spatial patterns in dual paths. For the temporal path, we introduce a perturbation positional encoding to improve the issue of signal misalignment caused by individual differences. For the spatial path, we propose a correlation metric based on Gramian angular field similarity to establish a more specific whole-brain functional network. Subsequently, we interleave the query and key vectors of dual paths to interact spatial and temporal information. We further propose integrating the dual-path attention into a tensor that retains spatiotemporal dimensions and utilizing 2D convolution for feed-forward processing. Our attention layer allows the model to represent spatiotemporal correlations of signals at multiple scales to alleviate issues of information distortion and loss. Our STDCformer demonstrates competitive results compared to state-of-the-art methods on the ABIDE dataset. Additionally, we conducted interpretative analyses of the model to preliminarily discuss the potential physiological mechanisms of ASD. This work once again demonstrates the potential of deep learning technology in identifying ASD and developing neuroimaging biomarkers for ASD.
静息态功能磁共振成像（rs-fMRI）是一种广泛应用于自闭症谱系障碍（ASD）研究的非侵入性神经成像技术，为深入了解ASD潜在的生物学机制提供了初步见解。深度学习技术在rs-fMRI分析中已展现出巨大潜力。然而，准确区分健康对照组和ASD患者一直是一项长期挑战。在这方面，这项工作提出了一种具有时空模式双路径交叉注意力框架的模型，名为STDCformer，旨在提高ASD识别的准确性。STDCformer能够在深度明确交互时空信息的同时，保留时间特异性模式和空间特异性模式。STDCformer的嵌入层在双路径中嵌入时间和空间模式。对于时间路径，我们引入了扰动位置编码来改善个体差异导致的信号错位问题。对于空间路径，我们提出了一种基于格拉姆角场相似性的相关度量，以建立更具体的全脑功能网络。随后，我们交错双路径的查询和键向量以交互空间和时间信息。我们进一步提出将双路径注意力集成到一个保留时空维度的张量中，并利用二维卷积进行前馈处理。我们的注意力层允许模型在多个尺度上表示信号的时空相关性，以缓解信息失真和丢失的问题。在ABIDE数据集上，我们的STDCformer与现有最先进方法相比展示出了具有竞争力的结果。此外，我们对模型进行了解释性分析，以初步探讨ASD潜在的生理机制。这项工作再次证明了深度学习技术在识别ASD以及开发ASD神经成像生物标志物方面的潜力。
