# An Autism Spectrum Disorder Identification Method Based on 3D-CNN and Segmented Temporal Decision Network

## TL;DR

This paper introduces a new method for identifying autism spectrum disorder using brain scans, combining advanced machine learning techniques to better capture brain activity patterns.

## Contribution

The novel framework combines 3D-CNNs and segmented temporal decision networks to improve spatiotemporal feature extraction from 4D fMRI data for ASD classification.

## Key findings

- The proposed method achieved an average accuracy of 0.85 on the ABIDE dataset with 1035 subjects.
- The approach outperforms existing state-of-the-art methods in capturing spatiotemporal brain activity patterns.
- The method provides insights into the neurobiological mechanisms of ASD through improved classification performance.

## Abstract

(1) Background: Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by social communication deficits and repetitive behaviors. Functional MRI (fMRI) has been widely applied to investigate brain functional abnormalities associated with ASD, yet challenges remain due to complex data characteristics and limited spatiotemporal information capture. This study aims to improve the ability to capture spatiotemporal dynamics of brain activity by proposing an advanced framework. (2) Methods: This study proposes an ASD recognition method that combines 3D Convolutional Neural Networks (3D-CNNs) and segmented temporal decision networks. The method first uses the 3D-CNN to automatically extract high-dimensional spatial features directly from the raw 4D fMRI data. It then captures temporal dynamic properties through a designed segmented Long Short-Term Memory (LSTM) network. The concatenated spatiotemporal features are classified using Gradient Boosting Decision Trees (GBDTs), and finally, a voting mechanism is applied to determine whether the subject belongs to the ASD group based on the prediction results. This approach not only enhances the efficiency of spatiotemporal feature extraction but also improves the model’s ability to learn complex brain activity patterns. (3) Results: The proposed method was evaluated on the ABIDE dataset, which includes 1035 subjects from 17 different brain imaging centers. The experimental results demonstrate that our method outperforms existing state-of-the-art approaches, achieving an average accuracy of 0.85. (4) Conclusions: Our method provides a new solution for ASD classification by leveraging the spatiotemporal information of 4D fMRI data, achieving a significant improvement in classification performance. These results not only offer a new computational tool for ASD diagnosis but also provide important insights into understanding its neurobiological mechanisms.