Deep Learning based Identification of Genetic Syndromes using Facial Images

A disease that affects a small number of people is termed as a rare disease. Around 4% of the total global population is afflicted with the rare diseases. Even if the symptoms don't manifest right away, the majority of uncommon diseases have a hereditary foundation and are therefore present for the entirety of the person's life. Genetic analysis, the ability to diagnose diseases is continually evolving, and testing is becoming more and more necessary. When sending patients to clinical genetics facilities, however, care must be taken due to the restricted resources, particularly the availability of trained clinical geneticists. Around 30–40% of genetic illnesses are linked to dysmorphic traits, which are unique face characteristics. Recent studies demonstrated the capacity of facial recognition technology to recognize genetic diseases. In this study, we evaluated how well deep learning face recognition model-based classifiers performed in identifying dysmorphic characteristics. Using facial photos and deep convolutional neural networks, this project enhances the early detection of uncommon craniofacial genetic syndromes including Down syndrome, Progeria, Apert and Williams’s syndrome. For the treatment of associated medical difficulties and health issues, early diagnosis of rare facial genetic syndromes is essential. According to studies, the existing screening procedures are insufficient, and in many cases, the disease is not discovered until a few years after the kid is born. We evaluated two classification problems, the first being a binary classification problem (Disease vs. Controls) and the second being a multi-class classification problem (5 genetic disorders vs. controls). The five genetic syndromes we are working on are: Down syndrome, Apert syndrome, Progeria Syndrome, Fragile syndrome, and Williams’s syndrome. We tried different methods for solving this problem and the best results we achieved were using the VGG-16 inspired CNN. We reached a high level of accuracy by using our dataset with 5 categories to fine-tune the VGG-16 model, with a peak accuracy of 92 percent and an F1-Score of 0.93.With a maximum accuracy of 92 percent and an F1-Score of 0.93, we successfully improved the VGG-16 model using our dataset with 5 categories