Computational Tools for Annotating, Segmenting, and Registering Premature Retinal Fundus Images

Retinopathy of Prematurity (ROP) is a proliferative retinal vascular disease affecting premature infants with lower birth weights, often leading to vision loss due to retinal detachment. Early diagnosis is crucial to prevent vessel proliferation and associated complications. However, manual diagnosis by ophthalmology professionals is challenging, prone to variability, and lacks standardization. On the other hand, machine learning models trained to perform ROP diagnosis suffer from interpretability issues. To address these challenges, this thesis gives the first steps towards a computational biomarker-based ROP diagnostic system. Recognizing the importance of segmenting the retinal anatomy, and in particular discriminating retinal blood vessels into arteries and veins, we first introduce Momaku, an annotation tool tailored for fundus images. Momaku has been developed and validated in collaboration with clinicians at Sant Joan de Déu Hospital, in Barcelona, who used the platform to annotate arteries and veins from a publicly available dataset of ROP images that only contained generic blood vessel annotations. In addition, the resulting database is used to train an artery/vein vessel segmentation model, which paves the way for the extraction of relevant diagnostic biomarkers. Finally, with an eye towards combining multiple images of the same newborn into a single image mosaic (a functionality of interest for paediatric ophthalmologists in this context), we also propose a method for assessing the quality of retinal image registration based on observing similarity on registered overlap regions, without requiring manually-annotated ground truth.