HYBRID TRANSFORMER-CNN MODELS FOR ENHANCED AUTISM SPECTRUM DISORDER CLASSIFICATION USING CLINICAL AND NEUROIMAGING DATA

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by a highly heterogeneous presentation, posing significant challenges for early diagnosis. The subtle differences between ASD and non-ASD individuals, especially during early developmental stages, make accurate classification difficult. However, early detection plays a crucial role in improving developmental outcomes through timely intervention, enabling affected children and families to access specialized therapies and support systems. This study explores the potential of using clinical data combined with deep learning techniques for automated ASD classification. We evaluated various deep learning models, including 3D CNN ResNet50, sequential CNN, 2D CNN combined with XGBoost, 2D CNN ResNet101, and Transformer-based architectures like the standard Transformer and Swin Transformer integrated with CNN. The incorporation of clinical parameters alongside neuroimaging features facilitated more nuanced pattern recognition associated with ASD. Conventional CNN models yielded moderate classification accuracy, ranging from 60% to 78%. Transformer-based models demonstrated superior performance, with Swin Transformer achieving the accuracy of 75%, highlighting their importance in capturing intricate patterns and relationships in the data. The Swin Transformer, or "Shifted Window Transformer," is a type of Vision Transformer (ViT) architecture designed for computer vision tasks. It introduces a hierarchical structure with multi-scale feature representation, making it more efficient for image recognition tasks compared to traditional ViTs. The results show that hybrid models, specifically the Hybrid CNN+Swin Transformer, outperform both traditional CNN architectures and pure transformer-based methods, achieving the maximum classification accuracy at 80%. This implies that a more thorough method of identifying ASD-related patterns in brain imaging data can be achieved by fusing the global contextual understanding of the Swin Transformer with CNN's spatial feature extraction capabilities. These findings underscore the potential of using Transformer-based architectures in ASD classification, leveraging clinical data to improve precision in early detection. This research provides a foundation for future investigations into hybrid approaches that integrate multiple data sources, advancing automated diagnostic systems for neurodevelopmental disorders.