Application of Artificial Intelligence in Diagnosing Autism Spectrum Disorder based on Structural Brain MRI using EfficientNet-B0 and Convolutional Block Attention Model

Abstract

Background: Autism spectrum disorder is a progressive neurological disorder that affects 1 in 60 children; early detection is vital for disease control and to prevent further progression. The latest diagnostic criteria for autism spectrum disorder, based on assessing behavior, are often time-consuming, costly, and necessitate a specialist.

Aim: This study proposes a model that helps clinicians diagnose autism spectrum disorder based on MRI neuroimaging.

Subjects and Methods: This retrospective diagnostic modeling study proposes a deep learning framework that combines EfficientNet-B0 with a Convolutional Block Attention Model and advanced data augmentation strategies for diagnosing autism spectrum disorder. The model was trained on the mid-slice axial T1-weighted MRI scan that belongs to the ABIDE II KKI_1 dataset.

Results: The experimental results of the proposed model achieve an accuracy of 95.37%, with recall, precision, and F1-score all exceeding 93%, with better performance compared with traditional architectures such as SqueezeNet and ResNet50.

Conclusions: The study approach achieves a high level of accuracy, sensitivity, and specificity; thus, it may be viewed as a valuable computer-aided diagnostic tool in autism spectrum disorder detection.