Development of an R Package for Shiny App Deployment in the Cloud

Keywords: Deep Learning; Clinical Biomarkers; Retinal Vasculature; Retinopathy of Prematurity; Artery-Vein Segmentation
 
Summary: Quantitative retinal vascular analysis based on color fundus images relies on artery/vein (A/V) segmentation. It enables the extraction of clinically meaningful vascular biomarkers, including arteriovenous caliber relationships, tortuosity, branching and junction geometry, or measures of global vascular complexity. These reproducible and objective vascular measurements support, in turn, large-scale population analyses. Recently, we curated ∼1,200 publicly available A/V annotations to develop OCULAR-Net, a model trained by focusing on clinically relevant vascular regions and the preservation of biomarkers, outperforming state-of-the-art models in clinical biomarker fidelity.
However, two main observations arose from this study: 1) resulting segmentations have critical failure modes (e.g., vessel continuity, label swaps, optic disc uncertainty); 2) vasculature biomarkers are fragile; different formulations of the same biomarker may not agree with each other, and small alterations of the vascular tree lead to significantly different results.
The proposed thesis will address these two issues by simultaneously exploring alternative deep segmentation methods, post-processing for automatic refinement and the sensitivity and clinical grounding of different biomarker measures, especially vessel tortuosity and dilation.
The study’s aim is objective Plus disease quantification in retinopathy of prematurity, in close collaboration with Hospital Sant Joan de Déu and the Department of Ophthalmology. Therefore, students will benefit from direct guidance from reference ophthalmologists (Dra. Marta Morales, Dra. Alicia Serra), as well as computer vision experts (Pr. Oscar Camara, Dr. Adrian Galdran) throughout the thesis.
 
 
Supervisors: Oscar Camara, Gonzalo Plaza