Computerized quantificiation of facial weakness in facioscapulohumeral muscular dystrophy

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
Facioscapulohumeral muscular dystrophy (FSHD) is a rare hereditary and progressive muscular disease. One of the first and most characteristic symptoms of FSHD is asymmetrical weakness of the facial muscles. This weakness varies from minimal asymmetry to a complete lack of facial expression. Due to this weakness, patients are limited in the use of their facial muscles and are thus less able to express themselves in a social context, which can hinder social communication. However, at this moment studies on (the progression of) facial weakness and the consequences on communication are lacking and there is no validated outcome measure for facial weakness. Facial weakness is difficult to objectify and even more difficult to follow up over time. To facilitate future research, a standardized quantitative outcome measure for facial weakness in FSHD is required. Within this project a grading system for objectively measuring facial weakness and a diagnosis system for predicting FSHD from facial weakness were developed. A novel dataset was created consisting of facial video recordings of FSHD patients and healthy controls while performing various tasks. Video frames at rest and maximal expression were identified and manually labeled with 68 facial landmarks. Experts on facial weakness graded the video recordings of the participants on degree of facial weakness and assessed if FSHD was present. After extracting various types of facial features which were reported to quantify facial weakness, several machine learning systems were trained and evaluated using a newly developed system evaluation pipeline. Subsequently, the best systems were compared with human experts on agreement. The results show that the developed facial weakness grading systems perform in high agreement with the established ground truth, but that the agreement among human experts should be improved. Furthermore, the developed systems predicting whether a participant has FSHD perform above expert-level. It was found that combining multiple feature types gave the best results and that combining 2D and 3D features yielded better results than only 2D or only 3D features. It was also found that subtraction features were the most unreliable, although this is thought be related to insufficient head stabilization. Furthermore, the system evaluation pipeline provides a useful framework to further investigate the contribution of features to grade facial weakness and diagnose FSHD. The work in this thesis shows that it is possible to create an objective facial weakness grading system for FSHD patients with comparable performance to the current golden standard. Many research projects on the effect of various treatments on facial weakness within FSHD could benefit from an objective measure for reporting facial weakness. However, the current work should be improved in many ways before it can serve as such an measure, for example the number of participants should be increased, and a method for automatic landmark localization should be incorporated. The presented work provides a promising starting point, which could eventually lead to the development of a computerized standard for grading facial weakness within FSHD patients.
Faculteit der Sociale Wetenschappen