Towards Unsupervised Domain Adaptation for Diabetic Retinopathy Detection in the Tromsø Eye Study

Abstract

Diabetic retinopathy (DR) is an eye disease which affects a third of the diabetic population. It is a preventable disease, but requires early detection for efficient treatment. While there has been increasing interest in applying deep learning techniques for DR detection in order to aid practitioners make more accurate diagnosis, these efforts are mainly focused on datasets that have been collected or created with ML in mind. In this thesis, however, we take a look at two particular datasets that have been collected at the University Hospital of North-Norway - UNN.

These datasets have inherent problems that motivate the methodological choices in this work such as a variable number of input images and domain shift.

We therefore contribute a multi-stream model for DR classification. The multi-stream model can model dependency across different images, can take in a variable of input of any size, is general in its detection such that the image processing is equal no matter which stream the image enters, and is compatible with the domain adaptation method ADDA, but we argue the model is compatible with many other methods. As a remedy for these problems, we propose a multi-stream deep learning architecture that is uniquely tailored to these datasets and illustrate how domain adaptation might be utilized within the framework to learn efficiently in the presence of domain shift. Our experiments demonstrates the models properties empirically, and shows it can deal with each of the presented problems. The model this paper contributes is a first step towards DR detection from these local datasets and, in the bigger picture, similar datasets worldwide.