Explainable deep graph attention learning with enhanced residual SwinV2 transformer for interpretable facial paralysis severity assessment

Facial paralysis is a severely disabling disorder that causes weakness or damage to facial muscles, thus leading to abnormal or asymmetric facial movements. Timely and precise detection and grading of facial paralysis is more essential for appropriate treatments and rehabilitations. Conventionally, facial image analysis was a manual process relying on expert proficiency. Recently, deep learning (DL) techniques have revolutionized the field, especially in the context of facial and clinical image analyses. While DL enables automatic, accurate facial recognition, its black-box nature restricts its use in healthcare. Therefore, explainable artificial intelligence (XAI) is integrated to enhance transparency that provides visual evidence for medical decision-making. This study introduces an Explainable Deep Graph Attention Learning Framework for Facial Paralysis Identification (XDGAL-FPI). The proposed model is designed to automatically identify the severity of facial paralysis using facial image data. Initially, the input images are preprocessed, including face detection and alignment, color space conversion and contrast enhancement, noise reduction, normalization, and data augmentation to improve the quality and consistency of the input data. Besides, deep feature representations are extracted using an Enhanced Residual SwinV2 Transformer. The extracted features are then provided to a Heterogeneous Graph Attention Network that models the relational dependencies between multiple facial regions, such as the eyes, eyebrows, and mouth, for precise severity classification. Additionally, to improve the training efficiency and convergence behavior of the model, the Ranger optimizer is employed. Furthermore, the GradCAM + + technique is integrated to improve model interpretability by highlighting the important facial regions that contribute to the model’s prediction. The effectiveness of the proposed model is assessed using YouTube Facial Paralysis (YFP) Updated Dataset, which contains 14, 389 images categorized into four severity levels across different facial regions. The experimental results indicate that the proposed XDGAL-FPI model achieves improved performance compared with existing approaches across various evaluation metrics. Thus, the proposed XDGAL-FPI model can serve as a reliable and interpretable tool for automated facial paralysis severity assessment. The source code of this study is available at https://github.com/researcher010-debug/XDGAL-FPI .