A multi-view convolutional neural network method combining attention mechanism for diagnosing autism spectrum...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Autism Spectrum Disorder (ASD) diagnosis relies on subjective behavioral assessments.
• Current diagnostic models for ASD lack sufficient accuracy.
• Functional magnetic resonance imaging (fMRI) offers objective brain activity measurement but requires improved analytical models.

Purpose of the Study:

• To develop an accurate and objective method for Autism Spectrum Disorder (ASD) identification using multi-scale fMRI data.
• To propose MAACNN, a hybrid deep learning model combining unsupervised and supervised learning with attention mechanisms.

Main Methods:

• Utilized stacked denoising autoencoders for unsupervised feature extraction from multi-scale fMRI data.
• Employed multi-view convolutional neural networks (CNNs) for supervised classification.
• Implemented an attention fusion mechanism for multi-scale data integration.

Main Results:

• MAACNN achieved 75.12% accuracy and 0.79 AUC on the ABIDE-I dataset.
• The model demonstrated 72.88% accuracy and 0.76 AUC on the ABIDE-II dataset.
• The proposed method significantly outperformed existing ASD diagnostic models.

Conclusions:

• MAACNN offers a robust and accurate approach for ASD diagnosis using fMRI.
• The integration of unsupervised and supervised learning with attention mechanisms enhances diagnostic performance.
• This method represents a significant advancement in the objective clinical diagnosis of ASD.