Noninvasive anemia detection and hemoglobin estimation from retinal images using deep learning: a scalable solution for resource-limited settings

AI Generated Summary*

Three CNN architectures were trained to flag anemia and predict hemoglobin from 45-degree fundus photographs of 2265 South Indian participants with diabetes aged 40 and above, with age and sex added as inputs for the classification task and an 80/20 split at patient level. InceptionV3 led, with an AUC of 0.98, accuracy of 98%, sensitivity of 99%, specificity of 97%, and mean absolute error of 0.58 g/dL, although Bland-Altman analysis showed proportional bias. Against 255 external ultra-widefield images accuracy was 0.83, improving to 0.86 once cropped to 45 degrees. Saliency maps clustered near the optic disc; separate vessel analysis found higher tortuosity in anemic participants in both zones, lower density only in the outer zone, and mixed thickness findings.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.