DCOAT-Net: A Multi-Frequency with Dynamic Connectivity Attention based Network for ASD Classification and Cognitive Assessment
作者
Gokul M,Maheshkumar T,Lingesh S,R. Santhosh
出处
期刊:2021 Third International Conference on Inventive Research in Computing Applications (ICIRCA)日期:2026-06-03卷期号:: 1593-1603
标识
DOI:10.1109/icirca69024.2026.11570462
摘要
Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition in which the precise classification and cognitive assessment remain challenging; nevertheless, fMRI and sMRI progressions recover understanding of ASD’s neural mechanisms and analytic potential. Although multimodal neuroimaging methods exist, numerous ones rely on feature concatenation instead of deep integration, important to suboptimal fusion. Additionally, a major inadequacy is the lack of frequency-domain examination, as most approaches fail to differentiate between high-frequency and low-frequency mechanisms, both of which are indispensable for classifying ASD-related forms. To discourse these challenges, Proposed model suggested DCOAT-Net, a novel deep learning (DL) technique for ASD classification and cognitive assessment. After preprocessing, images pass into the Bimodal Feature Learner Network (BFLN), which captures structural structures from sMRI and functional dynamics from fMRI with separate backbones to avoid feature interference. Then the Multi-Frequency Feature Separation Unit (MFFSU) processes these features via (1) Cross-Modal Feature Embedding Layer (CMFE) intended for multimodal representation, (2) Local Context Extractor (LCE) for local details, and (3) Global Context Extractor (GCE) for global connectivity. The novel Dynamic Neural Connectivity Attention Unit (DNC-Att) examines temporal fMRI dependencies for ASD detection. The final features are attached via Adaptive Gated Multi-Modal Fusion, and the Multi-Functional Classification Head accomplishes ASD classification using SoftMax and cognitive assessment via regression for IQ and executive function assessment. In the end, the DCOAT-Net begins a new benchmark in automated ASD analysis, presenting a robust, explainable solution for neurodevelopmental analysis. The results show that our model beats all the existing models by attaining accuracy of 92.5%, precision of 94.8%, and recall of 95.5%. This designates that our outline has greater efficacy compared to state-of-the-art methods.