Down Syndrome Detection Based On Facial Features Using Deep Learning
Down syndrome is the most commonly occurring chromosomal condition. Patients with Down syndrome have an increased risk for heart defects, respiratory and hearing problems and the early detection of the syndrome is fundamental for managing the disease. Clinically, facial appearance is an important indicator in diagnosing Down syndrome and it paves the way for computer-aided diagnosis based on facial image analysis. In this study, we proposed a novel method to detect Down syndrome using photography for computer-assisted image-based facial dysmorphology. The dataset curated comprises normal images of different ethnicities along with down syndrome images. These images were augmented and trained on several deep learning models such as CNN, Xception, InceptionResnetV2, Resnet50. Out of these 4 models, the InceptionResnetV2 model gave the highest accuracy of 96.88%.
The dataset consists of a total of 106 facial images where 80 were identified to be Normal and the remaining 26 were identified to have Down Syndrome. The Down syndrome images were obtained from various sources on the internet. We collected nearly 130 normal images out of which 55 images were used by us for this work. This was done in order to maintain a comparable ratio of Down syndrome and normal images. The remaining normal images were obtained from several online sources. The graph regarding the gender is depicted in figure 1. There are a total of 56 women face images and 50 men face images. The dataset includes facial images from other nationalities such as India, America, Mongolia, China. The dataset includes various age groups such as 1 - 10, 10 - 20 up to 60 - 70
_jfif.jpg)
_jfif.jpg)
_jfif.jpg)
_jfif.jpg)
_jfif.jpg)
_jfif.jpg)
Results
The images were trained using 4 different Deep Neural Networks. The first Convolutional Neural Network, which was trained for 25 epochs, gave an accuracy of 86.36%. The second, third and fourth models were based on transfer learning using XceptionNet, InceptionResnetV2 and ResNet50 and were trained for 10, 10 and 25 epochs respectively. XceptionNet gave an accuracy of 95.45%. We were able to obtain an accuracy of 96.88% using InceptionResnetV2 and 50% using ResNet50.