Answers are grounded in this paper and cite their sources.
Early and reliable identification of autism spectrum disorder (ASD) remains challenging because conventional diagnostic procedures are labour-intensive and rely heavily on expert clinical evaluation. Automated facial image analysis has been proposed as a low-cost screening aid; however, existing studies often use raw images without explicit feature optimisation. This study proposes a four-stage framework for classifying facial images into ASD and non-ASD categories. The framework integrates Contrast Limited Adaptive Histogram Equalization (CLAHE) for contrast enhancement, Local Binary Patterns (LBP) for texture feature extraction, Grey Wolf Optimization (GWO) for selecting a compact subset of LBP features, and a DenseNet-169 convolutional neural network for classification. The proposed framework was evaluated using a publicly available facial-image dataset comprising 2,940 images (1,470 per class) obtained from Mendeley Data. Evaluation was performed using a held-out test set of 300 images. The model achieved an accuracy of 79.7%, a sensitivity of 86.0%, a specificity of 73.3%, and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.73. These findings indicate the potential of combining texture-based feature extraction, metaheuristic feature selection, and deep learning for computer-assisted ASD screening. However, the evaluation was limited to a single public dataset, employed image-level data splitting, and did not include external validation. Therefore, the proposed framework should be regarded as a preliminary screening approach rather than a clinically validated diagnostic tool.
No takes yet. Share an insight, caveat, or question.
Onyema et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: