Multiple facial image features-based recognition for the automatic diagnosis of turner syndrome

Because of the diversity of individual clinical symptoms and the lack of reliable diagnostic criteria based on the clinical features of appearance, the initial diagnosis of Turner syndrome (TS) mainly depends on the clinical characteristic of height, resulting in many patients with Turner syndrome being diagnosed with other diseases, such as dwarfism. To improve objectivity, reduce the burden on well-experienced endocrinologists, allow screening of suspected patients in under-developed areas and provide TS patients with early detection and early treatment, a facial image analysis-based computer-aided system for automatic face classification is proposed. The system is composed of facial image pre-processing, image feature extraction, and automatic classification. First, several unique appearance features are identified in different facial regions based on clinical observations by endocrinologists, including ocular distance, epicanthus, and the numbers and sizes of melanocytic nevi. Based on the characteristics, we trained a 68 feature-points face model. Then, distance between points, Gabor wavelet filtering and spot detection are applied to extract global features and local features, respectively, and Gabor features are reduced by principal component analysis (PCA). Finally, Support Vector Machine (SVM) and the Adaboost algorithm are used for classification. Although all subjects involved in this trial are Chinese, the method achieves an average accuracy of 84.6% on the training set and 83.4% on the testing set based on K-fold cross- validation. The sustainable acquisition and accessibility of face images used for research is one of our advantages. We believe that this work can serve as an important reference for other assistant diagnosis systems related to facial images.