Development of deep learning model for autism spectrum disorder diagnosis

Abstract

The study investigates the application of deep learning-based techniques for Autism Spectrum Disorder (ASD) diagnosis using facial image datasets, aiming to contribute innovative insights to scientific literature. The research focuses on developing a robust framework for ASD diagnosis utilizing Convolutional Neural Networks (CNNs) like Xception, MobileNetV2, and ResNet50V2. A key innovation lies in leveraging facial image features as potential biomarkers to distinguish between ASD and Normal Control (NC) children. This approach is supported by the capacity of deep learning to extract nuanced facial features imperceptible to human observation. The study employs transfer learning via both model-centric and data-centric approaches to analyze datasets, including the Kaggle and YTUIA datasets. Hyperparameter tuning on the Kaggle dataset with the Xception algorithm achieves optimal accuracy of 0.95, surpassing prior studies and establishing benchmark settings. Further employing the method on the YTUIA dataset enhances accuracy performance to 0.965 by ResNet50V2, encompassing a broader demographic range and enriching ASD research by YTUIA. Explainable AI reveals that Xception focuses on central facial regions for ASD diagnosis, while ResNet50V2 and MobileNetV2 rely on peripheral features. To validate findings, the study introduces the UIFID dataset, comprising 130 ASD and Normal control (NC) samples. Models trained on Kaggle and YTUIA exhibit validation accuracies ranging from 0.72 to 0.79 and 0.45 to 0.60, on UIFID respectively, emphasizing the challenge of cross-domain validation. Addressing these limitations, data-centric strategies incorporating pre-processing and augmentation achieve peak accuracies of 0.989 on Kaggle, though declining to 0.823 on UIFID, necessitating enhanced feature generalizability. Advanced deep learning methods, including active learning, federated learning, and ensemble learning, are employed to mitigate domain divergence. Active learning achieves accuracies of 0.80 and 0.773 on combined test set (accumulated from Kaggle and YTUIA) and UIFID datasets, respectively, demonstrating iterative model improvement. Federated learning achieves 0.95 accuracy on combined datasets and addresses ethical concerns, while ensemble learning surpasses all methods with 0.96 accuracy on combined datasets and 0.901 on unseen UIFID data using the Fifty Percent Priority Rule (FPPR) algorithm which enhances ensemble learning by prioritizing models based on validation accuracies.