Genetic disorder and facial phenotype
There are more than 7000 known genetic disorders, and the number of affected is estimated to be about 6-10% of the population1. Around 30 to 40% of genetic disorders have physical changes in the face and skull such as Down's syndrome or Fragile X syndrome. Therefore, the known facial phenotype of many genetic disorders is highly informative to clinical diagnosis. As everyone knows, a fast and accurate diagnosis of genetic disorders is essential to prevent potential health problems. For clinical geneticists and pediatricians, it needs a high degree of experience and expertise to diagnosing genetic disorders through facial phenotype. However, there have some dilemmas for this issue. First, recognition of non-classical presentation or ultra-rare genetic disorder is constrained by the individual human expert's prior experience. Second, some genetic disorders will be confused in the clinic because they have a few subtypes (more than one typical phenotype) or overlapping facial characteristics with other disorders, such as the Cornelia de Lange syndrome. Lastly, the difficulty of diagnosis will increase due to facial phenotype sometimes being wide spectrum or becoming more prominent by age, like mucopolysaccharidosis. Briefly, genetic disorder diagnosis through facial phenotype still has a challenge.
Automated facial analysis
The research of computer-aided recognition has long been dealing with facial analysis-related problems, especially for non-classical presentation or ultra-rare genetic disorders. In other words, using computerized systems as an aid or reference for clinicians is becoming increasingly important2-6. In recent years, Face2Gene (FDNA Inc., Boston MA, USA) has been a novel and widely used tool to detect facial phenotype and recognize dysmorphic features from two-dimensional (2D) frontal photographs2. The facial dysmorphism novel analysis (FDNA) technology in Face2Gene is called DeepGestalt, which builds on deep convolutional neural networks (DCNNs) and uses computer vision and deep learning algorithms. The high-level flow of DeepGestalt is as described by Yaron Gurovich et al.. First is preprocessing a new input image to achieve face detection, landmarks detection, and alignment and then cropping the input image into facial regions. Second is feeding each region into a DCNN and obtain a softmax vector which indicate its correspondence to each syndrome in the model. Third is aggregating and sorting the output vectors of all regional DCNNs to obtain the final ranked list of genetic disorders. The part for the DCNN architecture of DeepGestalt is as a follow-up. There are ten convolutional layers in the network, and all but the last one are followed by batch normalization and a rectified linear unit (ReLU). A pooling layer is applied after each pair of convolutional (CONV) layer (maximum pooling after the first four pairs and average pooling after the fifth pair). And then the CONV layers are followed by a full connected layer with dropout (0.5) and a softmax layer. Therefore, a sample heatmap appears after each pooling layer. Comparing the low-level features of the first layers and the high-level features of final layer, the latter can identify more complex features in the input image and have tended to emerge distinctive facial traits when identity-related features disappear. Currently, the DeepGestalt model is trained on a dataset of over 17,000 images covering more than 200 different genetic disorders curated through Face2Gene, a community-driven phenotyping platform.
The article identifying facial phenotypes of genetic disorders using deep learning provide the reliability of DeepGestalt to diagnose genetic disorders through automated facial analysis. The binary gestalt model distinguishes a specific disorder from a set of other disorders. For Cornelia de Lange syndrome, DeepGestalt achieves 96.88% accuracy, 95.67% sensitivity, and 100% specificity. For Angelman syndrome, DeepGestalt achieve 92% accuracy, 80% sensitivity, and 100% specificity. Compared with previous related study, both have more precise diagnosis ability. The specialized gestalt model is used to classify different genotypes of the same syndrome. Noonan syndrome with a gene mutation in PTPN11, SOS1, RAF1, RIT1, or KRAS is a model used for testing the performance of DeepGestalt. In this study, DeepGestalt is a truncated version and predicts only the five desired classes. The result is the top-1 accuracy of 64%, which is superior to the random chance of 20%, allowing geneticists to investigate phenotype-genotype correlations. Multi-class gestalt model is that DeepGestalt performs facial gestalt analysis at scale. DeepGestalt has a 90.6% top-10 accuracy on the clinical test set and 89.4% on the publication test set. The top-5 and top-1 accuracy for the clinical test set achieve 85.4% and 61.3%, respectively, and for the publications test set, 83.2% and 68.7%, respectively. Therefore, the clinical can reach better prioritization and diagnosis of genetic disorders through an automated facial analysis framework. Potentially, DeepGestalt adds considerable value to evaluating the facial phenotype of a genetic disorder in clinical genetics, molecular study, and research.
In 2022, based on the previous work, Hsieh et al. proposed GestaltMatcher6, which utilized DCNNs trained on patients' photos as an encoder to convert facial photos into feature vectors to form a Clinical Face Phenotype Space (CFPS). They then quantified the similarity among patients by the cosine distance of two vectors in CFPS. With this approach, the investigators can support the ultra-rare syndromes that lack images to be trained and push the supported syndromes into the next level (from 299 to 1,115 syndromes). GestaltMatcher can also identify novel disorders. Moreover, it contributes to the longstanding discussion about distinguishability within the nosology of genetic diseases. Currently, there are 11 novel disease genes under analysis, and four of them were submitted to the peer-review journal.
