Anemia, defined by a reduction in hemoglobin and erythrocytes per unit of blood volume, represents a global epidemic: it affects more than one-third of the world’s population and poses a particularly severe threat to the health of young children and pregnant women. Unfortunately, against the backdrop of this massive problem, traditional laboratory-based screening methods require financial and technical resources, which makes their effective implementation impossible in low-resource settings or remote regions.
Clinical assessment of conjunctival pallor offers a fast and non-invasive alternative, but its diagnostic accuracy is critically dependent on both the assessor’s qualification and the ambient lighting.
The Role of Artificial Intelligence in Diagnostics
Recent technological progress has shown that Deep Convolutional Neural Networks (CNNs) have significant potential to make conjunctival assessment maximally objective and independent of the observer’s subjectivity or environmental conditions.
How Do Deep Convolutional Neural Networks Work?
Deep Convolutional Neural Networks (CNNs) are a powerful form of Artificial Intelligence widely used for analyzing visual data. Their design simulates the human brain’s processing of visual information.
The working principle of CNNs is based on breaking down images into small constituent parts – so-called features – such as texture, edges, or shapes. The extraction of these features occurs through multi-stage processing:
Convolutional Layers are responsible for identifying local visual patterns.
Pooling Layers reduce the dimension (size) of the data while preserving important information.
Fully Connected Layers ultimately perform object classification based on the extracted features.
The term “Deep” implies stacking many such layers together, allowing the network to learn increasingly complex, hierarchical, and abstract features. This ability makes CNNs indispensable for tasks such as object recognition, the diagnosis of pathologies in medical images, and the identification of patterns that are difficult for humans to perceive.
Initially, relatively simple image processing techniques and threshold-based methods were used to recognize conjunctival pallor, meaning the computer searched for a specific color intensity threshold in the image. The diagnostic accuracy of such an approach ranged from 80% to 84%.
The latest research is already using Deep Convolutional Neural Networks (CNNs), which are much more powerful than older methods. Progress has been achieved by using ensemble learning (combining several AI models for a better result) and training on Augmented Data Sets (i.e., enriching and multiplying existing data through artificial means). These methods have increased accuracy to almost 90%. The most significant innovation is that these AI models not only detect the presence of anemia but also have the additional function of predicting hemoglobin levels, which is particularly valuable in pediatric cohorts.
A systematic review of this comprehensive work was conducted according to PRISMA guidelines.
Key Research Results
For the study, 218 conjunctival images were selected (from the Eyes-Defy-Anemia dataset), including 126 healthy patients and 92 patients with anemia. Data was divided into training, validation, and testing subgroups according to the World Health Organization’s (WHO) hemoglobin criteria ( for non-pregnant women and for men).
Through the Google Vertex AI Automated ML Platform, four Convolutional Neural Network (CNN) models were tested: two models on palpebral (eyelid) views, and two on full anatomical images of the eye.
AI models that were “trained” on full eye images (with broad anatomical context) showed noticeably better diagnostic indicators. One of the models achieved 0.899 AUC and 100% sensitivity on palpebral tests. Nevertheless, models “trained” only on cropped eyelid views recorded limited sensitivity and a low Cohen’s , indicating that their results are less reliable. Additionally, the score calculated by the model was directly correlated with the severity of anemia.
For non-invasive anemia diagnostics to become reliable and globally accessible, the standardization of the image acquisition process and the enrichment of databases with various ethnic groups are critically important.
Source: American Journal of Hematology
