Deep-learning-based hemoglobin concentration prediction and anemia screening using ultra-wide field fundus images

AI Generated Summary*

A multitask convolutional neural network was trained to estimate hemoglobin concentration and detect anemia from ultra-wide-field fundus images acquired at one Chinese hospital. Altogether 11,528 images from 3,211 patients with paired blood tests were split into training, validation and test sets; age and sex were added to the classification stage. On 1,730 test images, hemoglobin prediction had a mean absolute error of 0.83 g/dl and anemia screening an AUC of 0.93. A second model trained on centrally cropped images, mimicking a conventional field of view, performed worse (MAE 1.21 g/dl, AUC 0.86). Saliency maps highlighted the optic disc, retinal vasculature and peripheral retinal zones. Noted limitations include the cross-sectional design and unexamined anemia subtypes.

*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.