Optimization and evaluation of facial recognition models for Williams-Beuren syndrome

Williams-Beuren syndrome (WBS) is a rare genetic disorder characterized by special facial gestalt, delayed development, and supravalvular aortic stenosis or/and stenosis of the branches of the pulmonary artery. We aim to develop and optimize accurate models of facial recognition to assist in the diagnosis of WBS, and to evaluate their effectiveness by using both five-fold cross-validation and an external test set. We used a total of 954 images from 135 patients with WBS, 124 patients suffering from other genetic disorders, and 183 healthy children. The training set comprised 852 images of 104 WBS cases, 91 cases of other genetic disorders, and 145 healthy children from September 2017 to December 2021 at the Guangdong Provincial People’s Hospital. We constructed six binary classification models of facial recognition for WBS by using EfficientNet-b3, ResNet-50, VGG-16, VGG-16BN, VGG-19, and VGG-19BN. Transfer learning was used to pre-train the models, and each model was modified with a variable cosine learning rate. Each model was first evaluated by using five-fold cross-validation and then assessed on the external test set. The latter contained 102 images of 31 children suffering from WBS, 33 children with other genetic disorders, and 38 healthy children. To compare the capabilities of these models of recognition with those of human experts in terms of identifying cases of WBS, we recruited two pediatricians, a pediatric cardiologist, and a pediatric geneticist to identify the WBS patients based solely on their facial images. We constructed six models of facial recognition for diagnosing WBS using EfficientNet-b3, ResNet-50, VGG-16, VGG-16BN, VGG-19, and VGG-19BN. The model based on VGG-19BN achieved the best performance in terms of five-fold cross-validation, with an accuracy of 93.74% ± 3.18%, precision of 94.93% ± 4.53%, specificity of 96.10% ± 4.30%, and F1 score of 91.65% ± 4.28%, while the VGG-16BN model achieved the highest recall value of 91.63% ± 5.96%. The VGG-19BN model also achieved the best performance on the external test set, with an accuracy of 95.10%, precision of 100%, recall of 83.87%, specificity of 93.42%, and F1 score of 91.23%. The best performance by human experts on the external test set yielded values of accuracy, precision, recall, specificity, and F1 scores of 77.45%, 60.53%, 77.42%, 83.10%, and 66.67%, respectively. The F1 score of each human expert was lower than those of the EfficientNet-b3 (84.21%), ResNet-50 (74.51%), VGG-16