Automatisierte Detektion des Schweregrades von Fazialisparesen mittels Convolutional Neural Networks (CNNs)

Facial palsy (FP) after tumor resection in the cerebellopontine angle significantly impairs patients' quality of life. The long-term rehabilitative potential correlates significantly with the postoperative severity of the paresis. FP classification is usually based on the House Brackmann (HB) score, which has a high inter-rater variability. The increasing integration of AI-based techniques in clinical and scientific routine potentially offers an automated, examiner-independent classification based on photographs and videos. However, it is not yet clarified which type of facial image (e.g., during smiling or frowning) or which deep learning algorithm is best suited for this purpose.

Different convolutional neural networks (CNN) (i.e., VGG-16, Darknet53, GoogleNet, SqueezeNet) were trained and evaluated for automated determination of facial palsy severity (according to the HB score) using > 8000 photographs, which were collected in a standardized process before and after cerebellopontine angle surgery. Performance metrics including accuracy, precision and recall were utilized to assess the effectiveness of the models. The comparative analysis between CNN architectures and different sets of images (i.e., smiling, frowning) aimed to identify the most suitable CNN architecture and data basis for accurate and robust facial palsy classification.

Among the set of CNNs evaluated, the Darknet53 provided the highest resolution for classifying the FP. Accuracy values > 90% were achieved to differentiate between HB grades. Notably, a classification based on photographs containing facial expressions with "smiling" patients was more successful than for images with "closed eyes" or "frowning". Data augmentation to balance the imbalanced groups (i.e., HB I-VI) appeared to be essential for improving prediction accuracy.

Classification of facial palsies with CNNs can be performed examiner-independently with high accuracy, demonstrating the potential of these approaches for automated assessment of facial palsy severity.