Automated Down syndrome detection from facial images using a hybrid multi-stage deep learning framework with adaptive attention

Automated detection of Down syndrome from facial images requires models that are robust to variations in illumination, pose, resolution, and background conditions. To address the limited generalization ability and the reliance on single-path feature learning reported in previous studies, this work proposes a hybrid deep learning framework that integrates a multi-stage convolutional architecture with spatial attention modules and complementary handcrafted descriptors. The network employs sequential convolutional stages that progressively refine intermediate feature representations, while attention mechanisms emphasize discriminative craniofacial regions. In parallel, handcrafted morphological and texture descriptors, including Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP), are extracted to capture structural cues that deep networks may overlook, particularly under suboptimal imaging conditions. A lightweight fusion module aligns and integrates both feature representations to enable robust representation learning with minimal computational overhead. The proposed framework was evaluated on a binary facial image dataset comprising individuals with and without Down syndrome using a predefined 60/20/20 train–validation–test protocol to assess its diagnostic performance. Experimental results show that the proposed framework achieved an overall classification accuracy of 98.18%, with a precision of 99.18%, sensitivity of 99.05%, and specificity of 99.45%. Additional evaluation metrics, including the Cohen's kappa, Matthews correlation coefficient (MCC), F1-score, ROC analysis, convergence curves, and confusion matrices, further support the robustness and diagnostic capability of the proposed framework. Despite employing dual-path feature extraction, the framework maintains computational efficiency with approximately 12.9 million parameters, making it lighter than several commonly used deep baselines. Ablation studies further demonstrate that the attention module improves localization-driven discrimination, handcrafted descriptors enhance robustness to acquisition variability, and the multi-stage architecture stabilizes the training process. Overall, the results suggest that combining engineered feature priors with hierarchical deep representations provides a reliable and computationally efficient framework for automated Down syndrome detection from facial images.