DHCA-Net: A Novel Dual-Stream Hierarchical Channel Attention Network for Explainable Autism Spectrum Disorder...

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• Autism spectrum disorder (ASD) diagnosis relies on behavioral traits, often involving subjective assessments.
• Facial morphology analysis presents a potential non-invasive biomarker for objective ASD screening.
• Current diagnostic methods can be subjective and time-consuming.

Purpose of the Study:

• To introduce DHCA-Net, a novel deep learning framework for automated ASD detection from facial images.
• To enhance discriminative learning using advanced attention mechanisms and adaptive tuning.
• To improve the objectivity and efficiency of ASD clinical screening.

Main Methods:

• Developed DHCA-Net, incorporating hierarchical channel, temporal-depth convolutional, and inverted residual multi-core attention modules.
• Implemented an adaptive refinement step for denoising clinical features.
• Utilized a multi-head feature fusion (MHFF) mechanism for deep spatial and contextual resource exploitation.
• Trained and validated the model on a dataset of 2936 facial images.

Main Results:

• DHCA-Net achieved 93.7% classification accuracy, a 0.9365 F1 score, and a 0.9887 AUC.
• The model demonstrated superior performance compared to DenseNet, Xception, EfficientNet, and Swin-Transformer.
• Achieved an efficient average inference latency of 70.14 ms.

Conclusions:

• DHCA-Net offers a highly accurate and efficient deep learning approach for automated ASD detection.
• The explainable framework shows significant potential for scalable clinical screening applications.
• Future research should focus on generalizing the model across diverse demographic populations.