View Proposal

---

Proposer

Heba Elshimy

Title

Oculomics for Systemic Risk: Predicting Cardiovascular and Metabolic Markers from Retinal Fundus Photographs

Goal

To predict a systemic biomarker (cardiometabolic risk) from a single colour fundus photograph using a foundation-model backbone.

Description

Oculomics is the prediction of systemic disease from ocular images. Landmark work from Google Health showed CNNs could predict cardiovascular risk factors — age, smoking status, blood pressure — directly from fundus photographs, and the field has expanded to renal and neurodegenerative markers. The fundus offers a non-invasive window onto microvasculature, making it a cheap, scalable screening signal. The aim of this project is to develop and train a model to predict a systemic target from fundus images and rigorously audit its fairness. Contained question: can retinal images predict the chosen marker, and does performance degrade across acquisition device or patient subgroup? Pipeline: (1) use a retinal/vision foundation model (RETFound or DINOv2) as a frozen feature extractor, with a lightweight classifier head; (2) baseline against a standard CNN (e.g. ConvNeXt) trained end-to-end; (3) predict the systemic target such as a cardiometabolic disease (cardiovascular disease CVD or diabetes). Explainability via saliency/attention maps to show whether the model attends to vasculature.

Resources

Data (public): - BRSET on PhysioNet + mBRSET for a phone-camera generalisation test; pre-computed foundation-model embeddings for both are also on PhysioNet - China-Fundus-CIMT (Scientific Data, 2025): bilateral fundus images with carotid intima-media thickness measurements

Background

Url

External Link

Difficulty Level

Moderate

Ethical Approval

None

Number Of Students

1

Supervisor

Heba Elshimy

Keywords

machine learning, foundation models, omics, ai in healthcare, medical imaging

Degrees