Real-time non-invasive hemoglobin prediction using deep learning-enabled smartphone imaging.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate hemoglobin measurement is crucial for medical assessments like preoperative evaluations and blood loss monitoring.
• Traditional invasive methods are inconvenient and unsuitable for rapid, point-of-care testing.
• Existing complex models limit frequent testing in mobile medical settings.

Purpose of the Study:

• To develop a novel, compact, and efficient system for accurate hemoglobin level estimation.
• To leverage deep learning and smartphone technology for accessible medical assessments.
• To facilitate rapid and frequent hemoglobin testing, especially in mobile medical environments.

Main Methods:

• A smartphone application captured eye images for analysis by a deep neural network.
• The EGE-Unet model performed eyelid segmentation, and the DHA(C3AE) model predicted hemoglobin levels.
• Model performance was evaluated using metrics like MIOU, F1 Score, accuracy, MAE, MSE, RMSE, and R^2.

Main Results:

• The EGE-Unet model achieved high performance in eyelid segmentation (MIOU 0.78, F1 0.87, accuracy 0.97).
• The DHA(C3AE) model showed promising results for hemoglobin prediction (MAE 1.34, R^2 0.34).
• The system is compact (1.08 M) with low computational complexity (0.12 G FLOPs).

Conclusions:

• The developed system offers a cost-effective, swift, and accurate non-invasive method for hemoglobin estimation.
• It eliminates the need for supplementary devices, enhancing treatment planning and patient care.
• The system has significant potential for frequent and rapid hemoglobin testing in mobile medical settings.