Multi-Slice Generation sMRI and fMRI for Autism Spectrum Disorder Diagnosis Using 3D-CNN and Vision Transformers.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Autism spectrum disorder (ASD) lacks definitive diagnostic indicators, prompting research into neuroimaging biomarkers.
• Limited data availability in neuroimaging studies necessitates advanced analytical techniques like transfer learning.

Purpose of the Study:

• To investigate potential autism spectrum disorder (ASD) indicators using structural MRI (sMRI) and functional MRI (fMRI) data.
• To address data limitations by employing transfer learning and novel data augmentation techniques for ASD classification.

Main Methods:

• Proposed utilizing four vision transformers (ConvNeXT, MobileNet, Swin, ViT) with sMRI and a 3D-Convolutional Neural Network (3D-CNN) with both sMRI and fMRI.
• Implemented various data generation and slice extraction methods from 3D sMRI and 4D fMRI scans across axial, coronal, and sagittal planes.
• Evaluated models on the NYU dataset from the ABIDE repository, comparing against VGG and ResNet transfer learning baselines.

Main Results:

• The 3D-CNN model with 50 middle fMRI slices achieved a maximum accuracy of 0.8710 and an F1-score of 0.8261.
• Using whole fMRI slices (excluding brain extremities) improved classification with 0.8387 accuracy and 0.7727 F1-score.
• Transfer learning with the ConvNeXt model on sMRI outperformed other transformers, demonstrating the efficacy of proposed methods.

Conclusions:

• The study validates the effectiveness of multi-slice generation combined with 3D-CNN and transfer learning for ASD classification.
• fMRI analysis, particularly with 50 middle slices, shows significant promise for identifying ASD indicators.
• Advanced AI models and neuroimaging techniques offer a powerful approach to improving ASD diagnosis.