Title: Detection of Anaemia from Retinal Fundus Images via Deep Learning

Authors: Akinori Mitani, Abigail Huang, Subhashini Venugopalan, Greg S. Corrado, Lily Peng, Dale R. Webster, Naama Hammel, Yun Liu, and Avinash V. Varadarajan

Publication Source: Nature Biomedical Engineering (2019)

Anaemia affects over 1.6 billion people worldwide, yet conventional screening relies on invasive blood draws that require trained clinical staff and specialized laboratory equipment. To address this barrier, medical researchers are exploring non-invasive diagnostic tools that leverage artificial intelligence to identify systemic health conditions through routine eye examinations.

A landmark study published in Nature Biomedical Engineering investigated whether deep learning algorithms could accurately predict haemoglobin (Hb) concentration and screen for anaemia using standard color retinal fundus photographs.

What Was the Study Testing?

Researchers trained and evaluated an advanced deep neural network on a massive dataset from the UK Biobank, analyzing fundus photographs from over 57,000 individuals alongside basic demographic metadata.

The AI system was evaluated across several key parameters:

• Non-Invasive Prediction: Estimating continuous haemoglobin levels directly from retinal blood vessel morphology without drawing blood.
• Metadata Integration: Combining fundus images with basic non-invasive physical data (age, sex, and blood pressure) to enhance diagnostic precision.
• Diabetic Subgroup Screening: Assessing screening performance specifically in diabetic patients, who routinely undergo annual retinal fundus examinations.

Key Findings: Diagnostic Accuracy & Model Insights

The trial demonstrated that deep learning applied to retinal photographs offers a highly reliable method for non-invasive anaemia screening:

1. Robust Predictive Performance

• Haemoglobin Estimation: The combined AI model achieved a mean absolute error (MAE) of 0.63 g/dL (95% CI: 0.62–0.64) in predicting blood haemoglobin concentration.
• Anaemia Detection: The model achieved an Area Under the Curve (AUC) of 0.88 (95% CI: 0.86–0.89) for detecting general anaemia and an AUC of 0.95 (95% CI: 0.93–0.97) for moderate-to-severe anaemia.
• Diabetic Subgroup Consistency: Among diabetic participants (n = 539), the system maintained an impressive AUC of 0.89, proving its viability as an automated add-on during routine diabetic retinopathy screening.

2. Interpretability & Visual Saliency

• Optic Disc & Vascular Focus: Saliency map visualizations revealed that the neural network primarily focuses on fine anatomical structures around the optic disc and main retinal vessel branches.
• Biomarker Discovery: The model relies on subtle microvascular variations and vessel coloration differences that are imperceptible to the human eye.

What Does This Mean for Healthcare & Screening?

The ability to detect anaemia from fundus photographs transforms standard eye exams into powerful systemic health assessments. Because retinal photography is fast, painless, and already widely available in eye clinics and optical centers, incorporating AI screening could catch undiagnosed anaemia early without extra blood draws or patient discomfort.

This technology holds immense potential for opportunistic screening—enabling patients visiting an eye clinic for routine checkups to receive immediate insights regarding their blood health.

The Bottom Line

AI-driven retinal analysis offers a promising, painless alternative to conventional blood tests for anaemia screening. While further prospective validation in broader global populations is ongoing, deep learning model integration into clinical workflows represents a major leap forward in non-invasive preventive care.

Always consult a qualified healthcare provider or ophthalmologist to discuss comprehensive screening options and blood health evaluations.

Read the full study in Nature Biomedical Engineering.