Deep image registration for composing spectral retinal images

Abstract

Eye diseases are very common among the population, which directly affect the quality of life. The anatomy of the eye is the complex structure that determines how the tissues will interact with light. Optical filters transmitting light at specific wavelengths can be used to obtain multispectral images of the eye fundus. This allows to observe various details of the organ such as blood vessels, nerves, and the retina. The issue of image registration is relevant since the combination of several color channels allows to expand useful information of one image which makes it possible to to improve the diagnosis of eye diseases. Due to a significant increase in computing power, availability of data, and methodological advances artificial neural networks have become the main tool for analyzing medical images. This work studies a deep image registration method based on the U-net architecture that produces high-quality spectral retinal images, to evaluate the alignment with a subsequent comparison with the existing dual-bootstrap iterative closest point algorithm using basic statistical methods. The studied technique is able to compete with existing ones.