Non-invasive hemoglobin estimation from conjunctival images using deep learning

Hemoglobin, a crucial protein found in erythrocytes, transports oxygen throughout the body. Deviations from optimal hemoglobin levels in the blood are linked to medical conditions, serving as diagnostic markers for certain diseases. The hemoglobin level is usually measured invasively with different devices using the blood sample. In the physical interpretation, some signs are traditionally used. These signs are the palms, face, nail beds, pallor of the conjunctiva, and palmar wrinkles. Studies have shown that conjunctival pallor can yield more effective results in detecting anemia than the pallor of the palms or nail beds. This study is aimed to predict the hemoglobin level by deep learning method, non-invasive, cheap, fast, high accuracy, and without creating medical waste. In this context, conjunctival images and age, weight, height, gender, and hemoglobin values were collected from 388 people who donated blood to the Turkish Red Crescent. A dataset was generated by augmenting the gathered data with body mass index data. Within the scope of this investigation, the limits of agreement (LoA) value at a 95% confidence interval was computed to be 1.23 g/dL, while the bias was established as 0.26 g/dL. The mean absolute percentage error (MAPE) values were determined to be 3.4%, and the root mean squared error (RMSE) was calculated to be 0.68 g/dL. These findings exhibit a successful outcome compared to similar investigations, signifying that this non-invasive method can be employed for hemoglobin level estimation. Furthermore, the estimated hemoglobin levels could aid in diagnosing several hemoglobin-related ailments.