A Preliminary Study on Deep Learning Models for Facial Nerve Paralysis Diagnosis

Facial Nerve Paralysis (FNP) is a critical condition that affects facial muscle movement, necessitating accurate and early detection. This study evaluates and compares the performance of deep learning models, including Convolutional Neural Networks (CNN) and pre-trained architectures such as ResNet50 and EfficientNetB0, for FNP detection. The dataset was preprocessed and split into training and validation sets, with models trained using transfer learning and evaluated using key performance metrics. Among the models, EfficientNetB0 demonstrated superior accuracy ($94.12 \%$), F1-Score (0.94) and AUC (0.978), outperforming both CNN and ResNet50 in precision and recall. These findings suggest that EfficientNetB0 provides a more robust and reliable approach for FNP diagnosis, making it a potential candidate for clinical decision-support systems. Future work will focus on enhancing model generalizability using data augmentation and cross-validation techniques.