Artificial Intelligence Algorithms Detect Anemia from Fundus Images

Rehana Khan Khan

Credit: University of New South Wales

Automated detection of anemia from retinal fundus images using artificial intelligence (AI) may benefit patients undergoing routine retinal imaging, according to new data presented at the 2024 Association for Research in Vision and Ophthalmology (ARVO) Meeting.¹

The classification algorithm, trained and developed using three network architectures, revealed all had high sensitivity and specificity in detecting anemia from the fundus images, despite very slight differences in those findings.

“The convolutional neural networks tested in this study hold promise for early and non-invasive diagnosis of anemia, thereby reducing associated health risks,” wrote the investigative team, led by Rehana Khan Khan, a PhD candidate at the University of New South Wales.

Anemia impacts nearly one-quarter of the global population, primarily women and children, with iron deficiency being its primary cause.² Effective treatments are available for anemia once early detected, improving quality of life, and alleviating the symptoms of iron deficiency.

However, early detection and monitoring is often a struggle, given the invasiveness of diagnostic tests, and the financial burden associated with screening.³ Improving screening methods in low- and middle-income countries where anemia is most prevalent is particularly important.

For this study, Khan and colleagues sought to evaluate the through machine-learning algorithms trained on retinal fundus images, a potential cost-effective screening method.¹ The data set comprised 2,265 participants aged ≥40 enrolled in an observational study in India. Investigators obtained both ocular and systemic clinical parameters and dilated 45-degree 4-field retinal fundus images, for each patient in the data set.

The team also extracted results from a biochemical investigation, including blood lipid levels, complete blood count, hemoglobin, and hemoglobin A1c (HbA1c). Patients in the data set were randomly separated into an 80% development set, divided into 70% training and 10% tuning, and a 20% validation set.

Khan and colleagues trained the classification algorithm on the Visual Geometry Group (VGG16), Inception V3, and ResNet50 architectures. The sensitivity, specificity, and accuracy were measured, and compared with clinical anemia data. In the analysis, receiver operating characteristic (ROC) curves were created to assess the area under the curve (AUC).

The InceptionV3 architecture displayed an AUC of 0.98 for anemia detection using only fundus images, showing 98.6% accuracy, 99.4% sensitivity, and 97.8% specificity. Furthermore, both the VGG16 and ResNet50 models exhibited an accuracy of 97.8%, a sensitivity of 99.9%, a specificity of 95.6%, and an AUC of 0.97.

After including the metadata and a complete blood count, the combined models predicted hemoglobin concentration, with a mean absolute error of 0.96 (95% CI, 0.94 - 0.98) and an AUC of 0.99. According to the analysis, the fundus features were predominantly concentrated on the spatial characteristics around the optic disc.

Khan and colleagues indicated the convolutional neural networks tested in this study were promising for early and non-invasive anemia diagnoses and could reduce anemia-related health risks. Importantly, they noted these successes could be useful for widespread screening.

“The success of our algorithms highlights their potential for widespread screening, emphasizing the importance of capturing fine spatial features around the optic disc in the process,” Khan added. “This research could enhance early detection methods for anemia, promoting better health outcomes for individuals globally.”

References

1. _{Khan RK, Pathan AQM Sala Uddin, Lin SH, Kelleher P, Maseedupally V, Raman R, Roy M. Automated anemia detection from retinal fundus images using artificial intelligence. Poster presented at the Association for Research in Vision and Ophthalmology (ARVO) 2024 Meeting, May 5–9, 2024.}
2. _{Jimenez K, Kulnigg-Dabsch S, Gasche C. Management of Iron Deficiency Anemia. Gastroenterol Hepatol (N Y). 2015;11(4):241-250.}
3. _{An R, Huang Y, Man Y, et al. Emerging point-of-care technologies for anemia detection. Lab Chip. 2021;21(10):1843-1865. doi:10.1039/d0lc01235a}