Enhancing CNN Performance for Facial Nerve Paralysis Detection: Limited and Imbalanced Dataset Problem

This paper compares standard artificial neural networks (ANNs) and convolutional neural networks (CNNs) for automatically identifying facial nerve paralysis using a special set of images. Despite CNN outperforming ANN in robustness and accuracy, the dataset's imbalance and limited size constrained its performance. To address these issues, various improvement measures were employed, including training optimization, data augmentation, and class balance procedures. The upgraded CNN model achieved a 97% accuracy rate, 1.00 precision, and 0.92 recall for the patient class, demonstrating its strong diagnostic skills. The CNN performance in accurately identifying between normal and FNP cases was validated using classification metrics and a confusion matrix. These results show that CNNs are a useful tool for helping doctors diagnose facial nerve paralysis, particularly when there isn't much data available. Future studies will explore advanced deep learning architectures and extensive datasets to enhance the performance.