Non-invasive anemia detection from conjunctiva and sclera images using vision transformer with attention map explainability

Iron-deficiency anemia, a prevalent global health issue, traditionally requires invasive procedures for accurate diagnosis, such as a blood sample for measuring hemoglobin (Hgb) concentration. Nevertheless, this marker can be visually assessed by observing external anatomical elements, such as the eye's conjunctiva and sclera. These regions often appear paler in anemic individuals, providing a visual sign of potential anemia. In this work, a non-invasive approach for anemia detection utilizing sclera-conjunctival images is presented. Using the Vision Transformer (ViT) model with a transfer learning approach, robust classification of anemia/no anemia is achieved. This methodology not only focuses on classification accuracy but also incorporates an explainability technique to provide visual insights into the decision-making process of the model. Experimental results demonstrated high accuracy, where an overall accuracy of 98.47% is achieved. The ViT model's performance is compared against established machine learning and deep learning algorithms to evaluate its effectiveness in anemia detection. The analysis of the results indicates that the ViT model, with its ability to focus on relevant image features when analyzing the explainability results, offers a promising alternative for anemia detection, potentially reducing the need for invasive diagnostic procedures.