Clinically Applicable Deep Learning for Diagnosis and Referral in Retinal Disease

Type: paper Slug: clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease—hassabis Sources: clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease—hassabis Last updated: 2026-05-13

---

Summary

De Fauw, Ledsam, Romera-Paredes, Nikolov, Tomašev, and colleagues, with Keane, and Hassabis (2018) developed a deep learning system for analyzing retinal OCT scans to diagnose and triage patients with sight-threatening retinal diseases. Evaluated in a clinical setting at Moorfields Eye Hospital, the system achieved referral recommendations matching expert ophthalmologists in 94% of cases, demonstrating one of the first real-world deployments of AI-assisted diagnosis in healthcare.

Core content

Problem: Retinal diseases (age-related macular degeneration, diabetic eye disease, glaucoma) are leading causes of preventable blindness. OCT imaging is essential for diagnosis but requires specialist interpretation, creating bottlenecks in referral pathways (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Introduction).

Approach: A two-stage deep neural network architecture:

1. Segmentation network: Identifies anatomical features in OCT scans (retinal layers, fluid, lesions) (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Methods).
2. Classification network: Uses segmentation outputs alongside raw OCT data to produce triage recommendations (urgent, semi-urgent, routine, observation only) (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Methods).

Clinical evaluation:

• Tested on ~1,000 patients in a simulated clinical setting at Moorfields Eye Hospital (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Results).
• Referral decisions matched senior ophthalmologists in 94% of cases (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Results).
• Did not miss any urgent referrals that experts identified (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Results).
• Performance was robust across different OCT machine types and patient demographics (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Results).

Clinical applicability design:

• The system outputs explainable recommendations with highlighted image features, not just binary diagnoses (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Discussion).
• Designed to fit into existing clinical workflows rather than replacing clinicians (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease §Discussion).

Connections

• Theme: clinical-AI
• Project: DeepMind-general
• Collaborators: Jeffrey De Fauw (first), Joseph R. Ledsam, Bernardino Romera-Paredes, Stanislav Nikolov, Nenad Tomašev, Pearse Keane (Moorfields Eye Hospital)
• Era: deepmind-ascent
• Venue: venue—Nature-Medicine
• Related: paper—a-clinically-applicable-approach-to-continuous-prediction-of-future-acute-kidney-injury — both “clinically applicable” AI papers from DeepMind Health

Honest Gaps

• Metadata lists only 4 co-authors; the actual paper has ~12 authors.
• The extraction includes an embargo notice from the pre-print version — this is an artefact.
• The evaluation was a retrospective simulation, not a prospective clinical trial — real-world performance may differ.
• Tested only at Moorfields Eye Hospital (a world-leading specialist center) — generalization to community optometry settings is unproven.
• The “94% match” metric is somewhat generous — disagreements with experts may have been counted as “acceptable clinical variation.”
• No evidence that the system actually reduced waiting times or improved patient outcomes in practice.