Effect of different modalities of facial images for diagnosis of ASD by deep neural network.

This research aims to explore the potential of various facial image types in diagnosing Autism Spectrum Disorder (ASD) through the application of deep learning neural networks. It delves into how deep learning algorithms perform with different facial image modalities, especially 2D and 3D, while addressing specific challenges associated with each. The methodology includes training deep learning models on distinct datasets and conducting an in-depth analysis of their accuracy and performance metrics. Significantly, the ResNet50V2 model recorded a 96.9% accuracy rate on the 2D dataset, and the Xception model achieved an 84.4% accuracy rate on the 3D dataset. These findings emphasize the strong capability of deep learning neural networks in making accurate ASD diagnoses from facial images. Nonetheless, the research reveals a stronger proficiency in handling 2D over 3D images, suggesting a need for more comprehensive 3D dataset training to improve three-dimensional image processing. Through evaluating the efficacy of different image modalities, this investigation enriches the field's knowledge base, highlights the necessity for robust dataset development, and charts a course for future studies to advance the precision and practicality of ASD diagnosis via deep learning approaches. [ABSTRACT FROM AUTHOR]