Dataset of human skin and fingernails images for non-invasive haemoglobin level assessment

Abstract

Anaemia, a decrease in total concentration of haemoglobin (Hb) in blood, affects substantial percentage of the population worldwide. Currently, the gold standard for determining the Hb level is the invasive analysis of venous blood. Yet, more and more research groups demonstrate the possibility of non-invasive Hb assessment using white light imaging of tissue sites where Hb is the main chromophore, in particular, fingernails. Despite the promising declarations, non-invasive Hb assessment via RGB-imaging is still poorly used in practice. The main reason is the difficulty in establishing the true accuracy of the methods presented in different works since they are tested on private datasets collected under different experimental conditions. Here we present an open dataset containing RGB images of skin and fingernails for patients with a known level of Hb, thus providing a single benchmark for researchers and engineers in the field, aimed at fostering translation of non-invasive imaging methods to the bedside.

Fig. 1

(a) Exemplary images of hands presented in the dataset for patients with different total blood haemoglobin level (b). An example of segmented areas of nail plates and skin areas.

Fig. 2

(a) The number of male and female patients presented in the dataset. (b) The age distribution for patients of different gender in the dataset. (c) Distribution of the total Hb level estimated using the certified (invasive) procedure. (d) Boxplots of the total Hb level for male and female patients.

Fig. 3

Intensity distributions in R, G and B channels of segmented regions of the nails (a) and finger skin (b) calculated for the whole dataset.

Fig. 4

Intensity distributions of mean R, G, and B channels and correlated color temperature of the white reference standard in images of the dataset.

Fig. 5

Summary of the model-building pipeline for the non-invasive blood Hb level prediction from RGB images. (a) An exemplary image of patient’s hand with marked nails regions (red bounding boxes), skin (blue bounding boxes), and the region used to obtain reference values (green box). (b) An example of segmented images of nails and skin. The dotted line shows the subregions of segmented nails and skin images used to calculate the optical descriptors. (c) The histogram of the distribution of Hb levels in the original dataset and its kernel density estimation with Gaussian kernel used for dataset balancing. (d) Probability distribution for point of the original dataset to be sampled into a balanced subset. (e) Comparison of histograms of Hb distribution in the original dataset and balanced subsample. (f) Scatter plot for the Hb level predicted by the model from optical descriptors on the Hb level determined using a hematological analyzer for validation and test sets. Dashed line corresponds to the line with identity slope and zero intercept. (g) Bland-Altman’s plot of the dependence of the difference between the predicted and true Hb level on its half-sum. (h) The value of the root-mean-square prediction error of the model on the validation and test data sets.