2.1. Automation of the Laguna ONhE Method for Estimation of Haemoglobin in the Optic Nerve Head

The presence of haemoglobin in the tissue was estimated by Laguna ONhE with reference to the colour of the vessels. Haemoglobin mainly absorbs green radiation and reflects most of the red. Therefore, the reference colour of the vessels was calculated using the values of the red (R) and green (G) channels of their pixels, to which the formula (R−G)/R is applied. The same equation was used for the pixels of the tissue, and finally, the result was expressed as a percentage. An estimate of cup size and position was also obtained, and the results of the cup, rim sectors, vertical cup/disc ratio (Hb-C/D) and cup/disc area ratio were compared with the percentiles achieved in the normal population [10]. Each fundus camera model was calibrated to achieve an equivalent response.

To achieve full automation of the method, five neural networks were used: one to segment the edges of the optic nerve already described [13], one for vessel segmentation using 4195 optic disc images, one to identify the eye as left or right using 4201 images, one to recognize image quality using 7048 images, and one to identify between normality (using 1518 images) and confirmed or suspected glaucoma (using 1596 images). The technical method is described in detail in the Appendix A “Computing development setup”.

The classification results obtained by deep learning were associated with the distribution of haemoglobin and the estimated Cup/Disc ratios to define a new value for the “Globin Distribution Function” (GDF) index, as previously described [4,5]. Once the value of the deep learning classifier was normalized to the mean values and standard deviation in the normal population, it influenced the result of the current GDF index by 45%. In the remaining 55%, the rest of the usual variables that we used in previous studies also intervened with normalized values. An example of the graphic results is shown in Figure 1. It shows how the method automatically performs the analysis of a retinography (a). Once its quality has been checked, the position of the optic nerve is identified, the inner edge of Elschnig’s scleral ring is defined, which is more internal than the apparent clear edge, and the size and shape of the cup is estimated (b). Its veins and arteries are then segmented, the colour of which is used to estimate the relative haemoglobin of the rest of the nerve tissue, shown in a colour code (c). In each sector of the optic disc, cup or rim, their areas are estimated as a percentage of the total disk area and expressed in colour if it corresponds to what is expected in a normal optic nerve (d).

Example of Laguna ONhE analysis: (a) Original wide-field eye fundus image. (b) Identification of the optic disc and boundaries segmentation. Central cup estimated from the haemoglobin distribution. (c) Segmentation of reference vessels and pseudo-colour image of haemoglobin distribution. (d) Estimated sector areas as a percentage of the total area, and compared to a normal reference population.