A Dual Track Feature Fusion Network for ASD Detection using Swin Transformers and Convolutional Neural Network

Author:

Radhakrishnan Menaka1,Ramamurthy Karthik1,Shanmugam Saranya1,Prabhu Aneesh Jayan1,Sivaraman Balamadhan1,Janarthanam Aishwar Haris1

Affiliation:

1. Vellore Institute of Technology

Abstract

Abstract

Early diagnosis of Autism Spectrum Disorder (ASD) plays a crucial role in enriching a child's development, particularly in improving social communication, language development, and addressing behavioural challenges. Early signs of autism may be observable in childhood, but a formal diagnosis often occurs later in life. Behavioural-based assessments, such as the Autism Diagnostic Interview-Revised (ADI-R) and Autism Diagnostic Observation Schedule-Revised (ADOS-R), are currently used for diagnosing ASD. These methods of diagnosis are time-consuming and require trained professionals. Due to these disadvantages of the traditional method of diagnosis, deep learning is used, where feature extraction is done automatically from Magnetic Resonance Imaging (MRI) data, eliminating the reliance on subjective pre-defined features. This advancement not only captures subtle information that may be missed by human-defined features but also enhances accuracy significantly. The dataset comprises of axial view of MRI images from ABIDE-I dataset from Autism Brain Imaging Data Exchange (ABIDE) database. This study proposes a dual-track feature fusion network architecture comprising Swin Transformer and customised Convolutional Neural Network (CNN) for precise classification. Swin Transformers excel in capturing long-range dependencies within images, facilitating a deeper understanding of interrelations among different image components. Concurrently, CNNs are adept at extracting local features, thus contributing to improved classification performance by considering both local and global features. The experimental outcomes highlight the efficacy of the proposed feature fusion network, showcasing an accuracy rate of 98.7%, precision of 98.12%, recall of 98.77%, and an F1-score of 98.65% upon evaluation using the ABIDE dataset.

Publisher

Research Square Platform LLC

同舟云学术

1.学者识别学者识别

2.学术分析学术分析

3.人才评估人才评估

"同舟云学术"是以全球学者为主线,采集、加工和组织学术论文而形成的新型学术文献查询和分析系统,可以对全球学者进行文献检索和人才价值评估。用户可以通过关注某些学科领域的顶尖人物而持续追踪该领域的学科进展和研究前沿。经过近期的数据扩容,当前同舟云学术共收录了国内外主流学术期刊6万余种,收集的期刊论文及会议论文总量共计约1.5亿篇,并以每天添加12000余篇中外论文的速度递增。我们也可以为用户提供个性化、定制化的学者数据。欢迎来电咨询!咨询电话:010-8811{复制后删除}0370

www.globalauthorid.com

TOP

Copyright © 2019-2024 北京同舟云网络信息技术有限公司
京公网安备11010802033243号  京ICP备18003416号-3