Transforming Fundus Photography for Deep Learning-Based Anemia Screening

Background/Objectives: This study aimed to develop and evaluate a noninvasive anemia screening method using fundus photographs analyzed by deep learning models, by transforming circular fundus photographs into square images suitable for convolutional neural network analysis. Methods: A total of 39,036 fundus images and clinical data collected from 2011 to 2023 were used. Eligible patients had clearly visible macula and optic discs and corresponding hemoglobin measurements. Three image preprocessing methods were applied—untransformed, stretching, and remapping. Results: Among preprocessing methods, the stretching method yielded the most accurate predictions. Deep learning analysis was conducted using multi-tasked cascaded convolutional neural networks. The hemoglobin prediction model based on EfficientNet B5 with remapping preprocessing achieved good performance (R2 = 0.781, MAE = 1.23), while the anemia classification model demonstrated screening accuracy with AUC values exceeding 0.893. Conclusions: Deep learning analysis of fundus photographs demonstrates potential as a noninvasive screening method for anemia. This approach may be particularly beneficial for patients requiring regular hemoglobin monitoring.