Automated Detection of Facial Palsy from Facial Images Using Convolutional Neural Networks

Stroke is one of the leading causes of death and disability worldwide. If a diagnosis can be made quickly and accurately, then patients have a good chance of treatment with medication and the Mortality rate can be reduced. This study centers on detecting facial palsy of using “Face Images from Acute Stroke Patients and Non-Acute Stroke Patients” dataset, the dataset contains 3800 images which can be divided to into two groups, one group about patients with acute stroke and then other without. An 11-layer Convolutional Neural Network (CNN) model is built as part of this research to differentiate between these two categories. The dataset is divided into training and test sets, with 70% (2660 pictures) designed in training and 30% (1140 photos) for testing the efficiency. In the training phase the model performed impressively and achieved an accuracy of 98.86% with the learning rate of 5.0000e-04. The accuracy rate in the test phase is 97.52%, meaning this model is quite effective in identifying stroke patients. This result indicate that the CNN model is highly effective in identifying acute stroke from facial images. If early diagnosis can be made, and patient care quickly adjusted, it should be a promising approach for turning around this disease. Its high validation and test accuracy emphasis indicate the robustness of the model-and larger applications in future clinical settings.