Approved Research

How a Machine Looks Into the Eyes of a Human - Explaining how artificial intelligence deduces properties from retina images, to enrich the medical understanding of eyes and overcome regulatory barrier

Lay summary

AI models are able to see properties in the human retina that physicians didn't know were possible. For example, an AI model can deduce the self-reported sex by looking into their eyes, where humans do not know in what way the retina of men and women differ.

We aim to understand how to determine the different properties from the human retina, such as height, self-reported sex, blood pressure etc.. In particular, we want to analyze the prediction strategy of an AI that is able to predict these properties from images of the eyes. Our methodology is to explain machine learning models that are trained on the retina images, to make their prediction strategy fully transparent.

Such AI algorithms are very powerful in research and industry. In the medical domain there are strict regulations to use ML algorithms, since the prediction has to be transparent and traceable for the user. Explanation methods, particularly the 'layer-wise relevance propagation' method are able to tackle regulatory barriers. Which finally enables us to use these powerful algorithms in practice.

We aim to pursue the research within 3 years.

The potential impact on the public and private health sector is twofold. On the one side we aim to find what features in the human body or behavior influences the shape of the retina, that has not been known until now. This may lead us to rethink aspects of the medical understanding of the human eye. On the other side we will provide a methodology how such algorithms pass medical regulations. This may finally enable any public or private organization to make use of this technology, for example in biometric retina identification.