Female autism categorization using CNN based NeuroNet57 and ant colony optimization

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Autism identification and classification using medical image analysis is advancing.
• Autistic females exhibit distinct phenotypic and age-related brain variations compared to males.
• Limited female phenotypic and genotypic data hinders accurate diagnosis and manual assessment.

Purpose of the Study:

• To develop an advanced deep learning model for accurate medical diagnosis of autism spectrum disorder (ASD) in females.
• To address the deficiency in female-specific data for autism research.
• To improve the efficiency of diagnosing autistic females.

Main Methods:

• Proposed a 57-layer Convolutional Neural Network (CNN) architecture, NeuroNet57, for feature extraction from fMRI data.
• Pre-trained NeuroNet57 on a Brain Tumour dataset and then applied it to Autism Brain Imaging Data Exchange (ABIDE)-I+II datasets using T1 modality fMRI scans.
• Utilized ant colony optimization (ACO) for feature subset selection and dimensionality reduction, followed by classification using nine machine learning algorithms.

Main Results:

• NeuroNet57 extracted feature matrices of 14372 × 4096 for ABIDE_I and 16168 × 4096 for ABIDE_II.
• The K-Nearest Neighbors (KNN)-based fineKNN (FKNN) classifier achieved 92.21% accuracy on ABIDE-I and 93.49% on ABIDE_II.
• The proposed model demonstrated effectiveness in categorizing females with ASD from control subjects.

Conclusions:

• The NeuroNet57 model, combined with ACO and FKNN, shows significant promise for accurate and efficient identification of autism spectrum disorder in females.
• This research highlights the potential of deep learning in analyzing neuroimaging data for gender-specific autism diagnosis.
• Further research with larger, diverse datasets is warranted to validate and refine these findings.