A Comparative Study of Lightweight Uncertainty-Aware Bayesian CNNs and Ensemble Models for MRI-Based ASD Diagnosis

Abstract Background: Early and accurate diagnosis of Autism Spectrum Disorder (ASD) is essential for improving clinical outcomes. Although deep learning (DL) has demonstrated strong potential in analyzing structural Magnetic Resonance Imaging (sMRI), most existing frameworks rely on deterministic models that lack uncertainty quantification—an essential component for clinical decision-making—and often involve high computational complexity, limiting real-world applicability. Objective: This study aims to conduct a comprehensive comparative evaluation of three deep learning paradigms to develop a robust and uncertainty-aware framework for ASD diagnosis using multi-orientational MRI data (axial, coronal, and sagittal views). Methods: We systematically evaluated three categories of models: (1) fine-tuned deterministic architectures, including EfficientNetV2B0, EfficientNetB3, and MobileNetV2; (2) ensemble learning strategies, including averaging, majority voting, weighted averaging, and stacked generalization; and (3) lightweight Bayesian convolutional neural networks (CNNs) incorporating Monte Carlo (MC) Dropout and Variational Inference (VI) to estimate predictive uncertainty. This hierarchical framework enables a balanced assessment of diagnostic accuracy and model reliability. Results: Experimental results demonstrate that ensemble models achieved the highest classification performance, with a peak accuracy of 99.51% in the coronal view. Notably, the proposed lightweight Bayesian CNNs delivered competitive performance while providing well-calibrated uncertainty estimates, enabling the identification of ambiguous and out-of-distribution cases. Additionally, Grad-CAM visualizations revealed that model attention is concentrated in clinically relevant regions, particularly the frontal and temporal lobes, supporting the neurobiological plausibility of the predictions. Conclusion: This study highlights that integrating uncertainty quantification into lightweight deep learning architectures enhances both transparency and reliability compared to conventional deterministic approaches.