Machine/deep learning-assisted hemoglobin level prediction using palpebral conjunctival images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Palpebral conjunctival hue is a qualitative indicator for non-invasive anemia screening.
• Current methods for assessing anemia via conjunctival hue are subjective and lack quantitative precision.

Purpose of the Study:

• To develop and evaluate machine/deep learning models for quantitative hemoglobin prediction using smartphone-derived palpebral conjunctival images.
• To assess the diagnostic performance of these models for anemia detection.

Main Methods:

• Construction of a U-net segmentation model for isolating palpebral conjunctival regions.
• Development of non-convolutional neural network (CNN)-based and CNN-based regression models for hemoglobin prediction.
• Evaluation of model performance using correlation coefficients, sensitivity, and specificity.

Main Results:

• The CNN-based model achieved a higher correlation coefficient (0.44) compared to the non-CNN model (0.38).
• The CNN-based model demonstrated improved sensitivity (20%) and specificity (99%) for anemia detection over the non-CNN model (13% sensitivity, 98% specificity).
• Model performance was slightly enhanced by correcting for image aspect ratio and exposure time, with the lower conjunctiva proving crucial for prediction.

Conclusions:

• CNN-based deep learning models show superior performance for predicting hemoglobin values from palpebral conjunctival images compared to non-CNN models.
• Further improvements in prediction accuracy are anticipated with larger datasets, particularly including more cases of anemia.