*1.3. Related Works*

Over the years many researchers have put in effort toward developing non-invasive methods for anemia detection through hemoglobin estimation. The relevance of conjunctiva hue in the clinical evaluation of anemia was tested in [25] for 219 healthy ambulatory subjects. Three educated non-clinicians, appropriately trained, overall agreed on conjunctiva hue performing with kappa coefficients between 0.27 and 0.34. As a result, hue variation strictly depends on the objective of the assessment and training of field personnel. Comparing earlier results obtained by physical examination and the latest digital photography, the latter is minimizing variance, optimizing specificity and sensitivity by using machine learning and automatic segmentation procedures. Establishing the most successful technology still leaves questions about the best region to analyze exploiting color properties associated with better results. Studies in [26] from an ophthalmology point of view open a debate for correlation of anemia between bulbar conjunctival blood column and palpebral conjunctival hue (PCH). From the results of this study, it seems that the bulbar conjunctiva can be successfully included in the set of interesting features, achieving slightly less specificity than PCH, but higher sensitivity. Paradigms of non-invasive and on-demand diagnostics based on smartphone and digital images are spreading due to the advancing of remote diagnosis and affordability [27–29]. A smartphone camera-based application monitoring blood hemoglobin concentration has been developed in [30]. Utilizing a light source pointed to the patient's finger, they performed a chromatic analysis on 31 samples, achieving sensitivity and precision of 85.7% and 76.5% respectively; they received Food and Drug Administration agreement. Another smartphone-based self-screening tool is depicted in [31] utilizing fingernail beds digital images. Patients select the regions of interest by themselves, corresponding to the nailbeds, and a result is then displayed on the smartphone screen; camera flash reflections and white spots which may affect Hgb level measurements are removed with a quality control algorithm. They reported an accuracy of ±0.92 g/dL−<sup>1</sup> of CBC hemoglobin level with personalized calibration, suggesting the relevance of those systems as a monitoring utility. In our study, we analyzed assumptions from related past works and the clinical correlation between conjunctival pallor and anemia condition [32], proposing a fully automated segmentation algorithm. Throughout this process, color features from hemoglobin reflectance spectrum provide a key role in biasing towards a region of interest proposal.

In the literature, few works deal with the automatic segmentation of the conjunctiva. In particular, reference [33] proposes a method for the automatic segmentation of the palpebral conjunctiva that carries out an image processing process based on the equalization of the image in RGB, filter unsharp masking and red channel masking. In [34] the authors developed an algorithm for automatically segmenting the image by finding a "distinctly red" region, bounded by two parallel long-running edges at the top and the bottom; this is achieved by combining the Canny edge detection technique with morphological operations in the CIELAB color space. However, with the aim of estimating anemia, they stated that their method of segmenting was less reliable than manual conjunctiva segmentation made by an expert physician. In [35] the authors use a threshold triangle (which uses triangle algorithm for thresholding) for binary differentiation between the palpebral conjunctiva and background.

### *1.4. Image Capturing Methodology*

The technique adopted to capture digital images of a patient's conjunctiva was based on the latest approach of a research study conducted in [36–38]. As a recap, the main requirements to designing an effective tool for estimating the condition of anemia through the use of digital images of the palpebral conjunctiva would be:


The acquisition system is shown in Figure 2. It consists of a macro-lens assembled into a specially designed, 3D-printed lightened spacer Figure 2a and a typical smartphone as in the real-life application Figure 2b. The lens can take high-resolution images being attached to a smartphone (we used the Aukey PL-M1 25 mm 10x macro lens). The LED lights can be powered directly from the smartphone or a battery applied to the cover of a smartphone. The lens is fixed on the plastic cover of the smartphone: this device allows for obtaining high resolution images close to the eye, insensitive to the ambient lighting conditions.

The dataset used in the present study, which will be described later, has been created with a Samsung S6 smartphone.

**Figure 2.** (**a**) The acquisition device consists of a special spacer and a macro lens to acquire images with a high-resolution smartphone at close range; (**b**) the moment of the acquisition of an image of the conjunctiva.
