The diagnosis of Autism Spectrum Disorder (ASD) remains a significant clinical challenge due to the subjective, and time-consuming of traditional diagnostic methods, often leading to delayed intervention. Although deep learning (DL) techniques have shown promise in improving diagnostic accuracy, existing approaches predominantly rely on individual pretrained models and have yet to achieve the precision necessary for dependable clinical use. While deep learning has shown promise, the systematic and comprehensive comparative analysis of diverse ensemble learning techniques for ASD detection, which have the potential to significantly enhance model performance by integrating multiple complementary models, remains underexplored. To address this gap, we propose a robust and efficient DL-based framework for ASD detection through facial image analysis. We fine-tuned multiple pretrained models – EfficientNetV2B0, EfficientNetB3, and MobileNet – on ASD-specific datasets. To develop a highly accurate, robust, and clinically applicable DL framework for ASD detection, we systematically implemented and compared several ensemble strategies, aiming to identify the most effective approach for reliable diagnostic support. Experimental evaluations on a publicly available facial image dataset demonstrate that the proposed ensemble methods outperform single-model baselines, achieving superior classification performance. Moreover, the ensemble framework exhibits competitive results relative to existing ASD detection methods, indicating improved generalizability and clinical applicability.
2025-07
Thesis
2025-08-24
Explainable and Uncertainty-Aware Deep Learning Models for Clinical Decision Support in Autism Detection
The diagnosis of Autism Spectrum Disorder (ASD) remains a clinical challenge, as
traditional approaches are often time-consuming and can delay interventions. Deep
learning (DL) has shown promise in improving diagnostic accuracy; however, facial
image-based ASD detection largely relies on single pretrained models, while MRI
based studies are often constrained by heavily preprocessed data or single-site
acquisitions, limiting generalizability. Moreover, the complexity and opacity of DL
models increase the risk of misclassification in ambiguous cases, emphasizing the
need for uncertainty estimation to assess prediction reliability. To address these
gaps, this study proposes a robust DL framework for ASD detection, independently
analyzing facial and MRI data using ensemble strategies and uncertainty-aware
modeling. The study was conducted in three phases: (i) fine-tuning multiple
pretrained CNN models (EfficientNetV2B0, EfficientNetB3, and MobileNet) on
ASD-specific facial and MRI datasets; (ii) implementing and comparing ensemble
strategies—including average voting, max voting, weighted averaging, and stacked
generalization—to identify optimal configurations; (iii) developing lightweight
Bayesian CNNs (BCNNs) with Monte Carlo Dropout (MCD) and Variational
Inference (VI) for uncertainty estimation, with Grad-CAM applied to highlight
discriminative MRI regions. Experiments on Kaggle (facial) and ABIDE II (MRI)
datasets showed that ensembles consistently outperformed single models.
