Fraunhofer Center for Assistive Information and Communication Solutions – AICOSDescription:
According to the World Health Organization, cardiovascular diseases have been the leading causes of death between the years of 2000 and 2016. Additionally, diseases of the circulatory system take up a high percentage of government budgets on health care systems, which could also be reduced with early detection. Recent studies have shown that age, blood pressure and the size of the optical disc vessels are correlated. Other studies claim there is a relationship between retinal vessels and cardiovascular diseases.
Fraunhofer Portugal AICOS has been working on research and development of innovative solutions for handheld retinal fundus image acquisition and Diabetic Retinopathy computer-aided detection. Fraunhofer's solution uses the EyeFundusScope prototype, a portable retinograph accompanied by an Android application, capable of processing the images captured by the device. This thesis aims to expand the potential use of this prototype, in the scope of preventive medicine, for screening Cardiovascular Diseases Risk through the automatic analysis of retinal fundus images.
During this dissertation, a small fundus photographs dataset, annotated using the SCORE cardiovascular risk calculator, was produced. We proposed a retinal blood vessel enhancement pipeline for improvement of the contrast between the retinal background and these structures. The output of the previous step was used to train a MobileNet architecture, pretained on ImageNet, achieving a three-class classification of 58.82% for low, moderate and high cardiovascular risk. The final model was able to identify the presence of risk (moderate or high) with a 79.41% accuracy.
The final results are encouraging and lead to the conclusion that there is a direct relationship between retinal blood vessels and cardiovascular risk. The solution shows it is possible to obtain cardiovascular risk stratification from fundus images in a mobile setting.
Author: David Azevedo
Type: MSc thesis
Partner: Faculdade de Engenharia da Universidade do Porto
Year: 2019