Automatic Facial Paralysis Evaluation Augmented by a Cascaded Encoder Network Structure

Facial paralysis refers to a facial nerve disordering, with which people may lose the abilities to accurately control their facial muscles for certain facial performances. The diagnosis of such disordering is mainly based on the observation of patient's face in terms of the facial spatial information, such as facial asymmetry. Up to now, this area is still dominated by therapists' subjective examinations clinically. Therefore, automations for this task receive wide attentions in both academic and industrial fields. Recently, the deep learning based methods, the convolutional neural networks (CNNs) more specifically, demonstrate their competitive performance compared with traditional approaches in many areas. However, due to the lack of the structured/labelled facial paralysis data as training data, those deep learning based solutions are still not able to fully attach their superiorities to the facial paralysis evaluation tasks. Another essential aspect for automation in facial paralysis analysis is the facial spatial information extraction. Semantic segmentation is a better choice than traditional template-based facial landmark detection for analysing facial paralysis images, which contain faces in uncommon patterns. However, most existing semantic segmentation approaches are made for indoor or outdoor scene parsing. To this end, this paper presents a deep learning-based approach for automatic facial paralysis grading prediction. The proposed model utilizes a cascaded encoder structure, which explores the advantages of the facial semantic feature for facial spatial information extraction, and then benefits the facial paralysis assessment. A dual-stage cascaded training process is adopted to utilize a mixture of normal and paralysed faces as training data, which exports a well-trained deep neural network model for facial paralysis evaluation. Experiments are conducted in two aspects to demonstrate the performance of each components of the proposed model. Encouraging results are illustrated compared with several existing approaches in the related areas.