Moorfields and UCL release nearly 1 million eye images to power AI research

Moorfields Eye Hospital and University College London have published one of the largest ophthalmic imaging datasets ever made available for research. The CADMUS dataset holds 945,243 images tied to clinical records from 22,482 patients, collected at Moorfields between December 2019 and September 2024. The goal is to help researchers build AI tools and study conditions affecting the front of the eye.

The timing matters. Anterior segment conditions, which include cataracts, keratoconus, and related disorders, are among the leading causes of blindness worldwide. Yet fewer than 10% of existing ophthalmic imaging datasets cover them. CADMUS is designed to fill that gap.

Researchers can apply for access through INSIGHT, Moorfields' Eye and Oculomics Health Data Research Hub. The application process includes independent patient and public oversight and follows the internationally recognised "Five Safes" framework, which evaluates whether a project is safe across five dimensions: projects, people, data, settings, and outputs.

How will it work?

The dataset is built from multiple layers of clinical data collected during routine patient visits. Because it includes follow-up appointments over nearly five years, it also captures how conditions change over time, which is critical for studying disease progression.

The data includes:

• Raw DICOM images from the MS-39 anterior segment OCT tomographer, a device that produces detailed cross-sections of the eye's front structures
• Derived measurements such as keratometry values (corneal curvature), pachymetry (corneal thickness), and wavefront aberrometry (optical distortions)
• AI-generated classifier scores for keratoconus and related conditions
• Linked electronic health record data covering demographics, diagnoses, and clinical history

Early results from teams already using CADMUS are encouraging. Lead author Shafi Balal says the dataset has helped establish precision limits for measuring keratoconus progression. His team has also trained deep learning models on it, including one that can predict a patient's age and biological sex from anterior segment scans alone. That finding suggests routine clinical images contain biological information that human clinicians cannot detect visually.

Why does it matter?

Datasets like CADMUS are rare. Most AI tools in ophthalmology have been trained on retinal images, where large public datasets already exist. The front of the eye has been far less represented, which has slowed research and limited the AI tools available to clinicians treating conditions like cataracts.

Moorfields is hoping the dataset accelerates work in three specific areas:

• Earlier detection of anterior segment disease
• Prediction of surgical outcomes
• Development of AI diagnostic and monitoring tools

The dataset is published in Ophthalmology Science and is available via a formal data access request. Full details are at doi.org/10.1016/j.xops.2026.101203.

The context

CADMUS arrives at a moment when health systems across the UK and Europe are pushing hard to make clinical data more useful for research. Several parallel efforts are underway.

Cambridge University Hospitals NHS Foundation Trust recently welcomed the launch of Zenith, a supercomputer funded by the Department for Science, Industry and Technology. Zenith is designed to help researchers study health data at scale and support AI development across the NHS. The AI Centre for Value-Based Healthcare is working with Zenith to make sure the system is used responsibly.

At the European level, the UNITE project, funded by the EU's Regional Innovation Valley programme, recently selected three projects from an open call that drew more than 1,000 organisations and 19 proposals. The focus: making health data usable and interoperable across fragmented systems.

Closer to home, London-based social enterprise PPL has launched the London Neighbourhood Public Data Explorer, a free tool that pulls together population health data to help identify local needs across the capital.

The common thread running through all of these efforts is that health data, collected during routine care, has research value that is only beginning to be realised. CADMUS is a concrete example of what that looks like in practice: nearly a million images, linked to real patient records, made available under a rigorous governance framework for anyone with a legitimate research purpose.