Ju Xiangyang, Ayoub Ashraf, Morley Stephen
Medical Devices Unit, Department of Clinical Physics and Bioengineering, NHS Greater Glasgow and Clyde, Glasgow G3 8SJ, UK.
Dental School, MVLS College, University of Glasgow, Glasgow G12 8QQ, UK.
Sensors (Basel). 2025 May 22;25(11):3264. doi: 10.3390/s25113264.
The subjective assessment of facial paralysis relies on the expertise of clinicians; the main limitation is intra-observer and inter-observer reproducibility. In this paper, we proposed a deep learning approach combining point clouds of facial movements with expert consensus to objectively quantify the severity of facial paralysis. A dynamic 3D photogrammetry imaging system was used to capture the facial movements of five facial expressions. Point clouds of the face at rest and at maximum expressions were extracted. These were integrated with the experts grading of the severity of facial paralysis to train a PointNet network to quantify the severity of facial paralysis. The results showed an accuracy exceeding 95% for assessing facial paralysis.
面瘫的主观评估依赖于临床医生的专业知识；主要局限性在于观察者内和观察者间的可重复性。在本文中，我们提出了一种深度学习方法，将面部运动的点云与专家共识相结合，以客观地量化面瘫的严重程度。使用动态3D摄影测量成像系统捕捉五种面部表情的面部运动。提取静止和最大表情时的面部点云。将这些点云与面瘫严重程度的专家分级相结合，训练一个PointNet网络来量化面瘫的严重程度。结果表明，评估面瘫的准确率超过95%。
