فهرست نشریات علمی دسترسی آزاد ایران

Introduction: Early and accurate detection of depression remains a pressing clinicalchallenge, especially in resource-limited environments. Facial expression analysis hasemerged as a promising, non-invasive screening method, yet many existing approaches areeither computationally intensive or lack clinical interpretability.Methods: This study aims to develop a lightweight, explainable deep learning framework fordepression screening using static facial images, with a specific focus on clinical relevance anddiagnostic transparency.Methods: We propose a hybrid architecture that leverages fine-tuned convolutional featuresfrom ResNet-18, followed by classification with a decision tree optimized using Gini impurity.Facial images were sourced from a publicly available dataset comprising over 20,000 labeledsamples, representing diverse adult populations. Images were preprocessed using contrastenhancement and bilateral filtering to preserve subtle affective cues. The model was trainedand evaluated using stratified 5-fold cross-validation, with performance assessed via accuracy,precision, recall, F1-score, and confusion matrix analysis.Results: The proposed framework achieved an average classification accuracy of 91.4%,outperforming several baseline visual-only models. Importantly, the use of a fine-tuneddecision tree classifier yielded clear, interpretable diagnostic rules that aligned with clinicalpreferences. The model demonstrated robustness across folds and strong generalizability,requiring minimal computational resources. Comparative analysis further highlightedthe method’s balance between performance and interpretability, making it well-suited forintegration into clinical decision support systems.Conclusion: This study demonstrates potential in combining deep learning-based featureextraction with interpretable classifiers for mental health screening. The method offers apractical, explainable, deployable solution for early-stage depression detection using facialimagery.