• SANJU S ANAND
• SANJU S ANAND
• Dr. Shashidhar Kini

Autism spectrum disorder (ASD) is a heterogeneous and complex neurodevelopmental disorder that poses remarkable challenges for early diagnosis. In fact, owing to the lack of distinguishing features between ASD and non-ASD subjects based on behaviour other than early detection (i.e., during normal development), accurate classification is extremely challenging [4]. However, early detection is vital in improving developmental outcomes, as timely intervention allows children and their families better access to specially designed therapies and support systems. The history and status of neuroimaging in autism. In this study, we explore the effectiveness of CNN models and a hybrid CNN-RNN architecture for ASD classification based on neuroimaging data, utilizing the Automated Anatomical Labeling (AAL) map to enhance feature extraction from the neuroimaging data. One important challenge in this task is how to construct representative feature maps that can capture the complex patterns of brain regions associated with ASD. To improve image quality and highlight critical structural details, we applied super-resolution techniques using Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN), which effectively enhanced neuroimaging resolution, facilitating better feature extraction. Additionally, edge detection methods such as Canny, Sobel, and Laplacian filters were employed to emphasize important boundaries and texture information within brain images, further aiding the models in identifying subtle differences between ASD and non-ASD subjects. Different types of models were evaluated, including VGG16 2D CNN, VGG16 3D CNN, EfficientNetV2, Inception v3, ResNet50 2D CNN, and a hybrid CNN-RNN model. The traditional CNN models achieved accuracies ranging from 60% to 70% at best, but the hybrid CNN-RNN model, which combines spatial and temporal features, significantly outperformed the others with an accuracy of 98%. Therefore, hybrid deep learning architectures play a significant role in ASD classification. Using AAL mapping, ESRGAN-enhanced neuroimaging, edge detection methods, and deep learning techniques, this study suggests that a semi-automated approach may be beneficial for forensic precision in early diagnosis of ASD. Our findings provide a foundation for future research on neuroimaging-based diagnostics and support the development of more complex hybrid models for autism classification as a neurodevelopmental disorder.

Autism Spectrum Disorder (ASD), Deep Learning, Convolutional Neural Networks (CNN), Hybrid CNN-RNN Models, AAL Map (Automated Anatomical Labelling), Neuroimaging, ASD Classification, Early Detection, Feature Extraction,

"Comparative Study of Deep Learning Models for ASD Classification with ESRGAN-Enhanced Images and AAL-Based Feature Extraction", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.12, Issue 6, page no.a634-a653, June-2025, Available :http://www.jetir.org/papers/JETIR2506064.pdf

"Comparative Study of Deep Learning Models for ASD Classification with ESRGAN-Enhanced Images and AAL-Based Feature Extraction", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.12, Issue 6, page no. ppa634-a653, June-2025, Available at : http://www.jetir.org/papers/JETIR2506064.pdf