Transforming Fundus Photography for Deep Learning-Based Anemia Screening

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Anemia screening traditionally requires invasive blood tests.
• Developing noninvasive methods for anemia detection is crucial for patient care.
• Fundus photography offers a potential noninvasive data source.

Purpose of the Study:

• To develop and evaluate a noninvasive anemia screening method.
• To utilize deep learning models for analyzing fundus photographs.
• To transform fundus images for convolutional neural network compatibility.

Main Methods:

• Utilized 39,036 fundus images and clinical data (2011-2023).
• Applied three image preprocessing techniques: untransformed, stretching, and remapping.
• Employed multi-tasked cascaded convolutional neural networks, including EfficientNet B5.

Main Results:

• The stretching preprocessing method showed the highest accuracy.
• The hemoglobin prediction model (EfficientNet B5 with remapping) achieved R² = 0.781 and MAE = 1.23.
• The anemia classification model demonstrated high screening accuracy with AUC > 0.893.

Conclusions:

• Deep learning analysis of fundus photographs is a viable noninvasive anemia screening method.
• This technique holds potential for patients needing frequent hemoglobin monitoring.
• Further research can refine this approach for widespread clinical use.