Abstract
Purpose :
Anemia affects about a quarter of the world's population, with its detection often disregarded due to the invasive characteristics of diagnostic tests and the financial burden associated with screening. This research aims to assess the potential of identifying anemia through machine-learning algorithms trained using retinal fundus images.
Methods :
The dataset for this study included 2,265 participants aged 40 years and above, who were enrolled from an observational study conducted at Sankara Nethralaya in Chennai, India. Clinical parameters, both ocular and systemic, along with dilated 45-degree 4-field retinal fundus images, were extracted. Additionally, biochemical investigation results, including blood lipid levels, complete blood count, hemoglobin, and HbA1c, were derived. The dataset was randomly split into an 80% development set, further divided into 70% training and 10% tuning, and a 20% validation set. The classification algorithm was trained and developed using the Visual Geometry Group (VGG16), InceptionV3, and ResNet50 architectures. Sensitivity, specificity, and accuracy were calculated, comparing the results with clinical anemia data. Receiver operating characteristic (ROC) curves were then generated to assess the area under the curve (AUC). The identification of fundus features associated with anemia was also explored.
Results :
For anemia detection using only fundus images, the InceptionV3 achieved an AUC of 0.98, with 98.6% accuracy, 99.4% sensitivity, and 97.8% specificity. Both VGG16 and ResNet50 models showed an accuracy of 97.8%, a sensitivity of 99.9%, a specificity of 95.6%, and an AUC of 0.97. When incorporating metadata and a complete blood count, the combined models predicted hemoglobin concentration (in g/dL) with a mean absolute error of 0.96 (95% confidence interval: 0.94–0.98) and an AUC of 0.99. The fundus features are predominantly concentrated on the spatial characteristics around the optic disc.
Conclusions :
Even though there were very slight differences, all architectures showed high sensitivity and specificity in detecting anemia from the fundus images. The convolutional neural networks tested in this study hold promise for early and non-invasive diagnosis of anemia, thereby reducing associated health risks.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.