Face Image Based Automatic Diagnosis by Deep Neural Networks

Abstract

In this paper, we use ResNet based networks for the automatic diagnosis of the Turner Syndrome (TS) by facial images. The TS is a common sex chromosomal disorder, which is due to the total or partial absence or structural abnormality of the X chromosome. Nowadays, the diagnosis of the TS mainly depends on peripheral blood lymphocyte chromosome karyotype analysis, which is time consuming. For inexperienced doctors, it is difficult to diagnose the TS only based on facial features, and there may be missed and inaccurate diagnosis. In order to help the TS patients to get timely diagnosis, we design and train ResNet-based networks to recognize patients' facial features, and build an intelligent system for automatic diagnosis. We evaluate the performance of the ResNet-based networks by sensitivity, specificity, and accuracy. We increase the average sensitivity from 67.6% to 91.54%, average specificity from 87.9% to 98.52%, compared with the AdaBoost method with local features. In the future, we aim to set up the intelligent system on a smart-phone to achieve fast and convenient screening of the TS at an early stage.