Related Experiment Video
Updated: Oct 29, 2025
12:21
Probing the Brain in Autism Using fMRI and Diffusion Tensor Imaging
Published on: September 12, 2011
25.4K
DarkASDNet: Classification of ASD on Functional MRI Using Deep Neural Network.
Md Shale Ahammed1, Sijie Niu1, Md Rishad Ahmed2
1Shandong Provincial Key Laboratory of Network Based Intelligent Computing, University of Jinan, Jinan, China.
Frontiers in Neuroinformatics
|July 12, 2021
Summary
A new deep learning model, DarkASDNet, accurately identifies autism spectrum disorder (ASD) using functional MRI (fMRI) brain scans. This advanced method achieves high accuracy in classifying ASD from typical controls, offering a promising tool for diagnosis.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Biomedical Engineering
Background:
- Non-invasive brain scans like functional Magnetic Resonance Imaging (fMRI) are crucial for diagnosing neuropsychiatric disorders, including autism spectrum disorder (ASD).
- Analyzing fMRI data for ASD diagnosis is challenging due to dataset limitations and the complexity of brain connectivity.
- Existing machine and deep learning models for ASD prediction using fMRI data often struggle with performance metrics like accuracy, precision, and recall.
Purpose of the Study:
- To develop an advanced deep learning model for accurate binary classification of autism spectrum disorder (ASD) versus typical controls (TC) using 3D fMRI data.
- To overcome the limitations of previous models in handling performance metrics and feature extraction for ASD detection.
- To establish a novel benchmark for ASD classification through improved accuracy and comprehensive evaluation.
Main Methods:
- Pre-processing of 3D fMRI data included slice time correction and normalization.
- Introduction of a novel deep learning architecture, DarkASDNet, designed for hierarchical feature extraction from lower to higher levels.
- Binary classification task to differentiate between ASD and TC groups using the developed DarkASDNet model.
Main Results:
- The proposed DarkASDNet model achieved a state-of-the-art accuracy of 94.70% in classifying ASD versus TC.
- Performance was validated using comprehensive evaluation metrics including precision, recall, F1-score, ROC curve, and AUC score.
- Results on the ABIDE-I, NYU dataset demonstrated superior performance compared to existing literature benchmarks.
Conclusions:
- The DarkASDNet architecture offers a novel and effective approach for ASD classification using processed fMRI data.
- The model's high accuracy and robust evaluation metrics suggest its potential as a diagnostic or prognostic tool for ASD.
- This work sets a new benchmark for ASD detection using deep learning on neuroimaging data.
