Development of an automated morphometric analysis processα-cell area was quantified on glucagon-stained slides (4 slides by patient, 2 from the head and 2 from the tail), β-cell area on insulin-stained slides (2 slides by patient, 1 from the head and from the tail), and non-exocrine-non-endocrine (non-acinar, non-insular) pancreatic tissue area on both insulin- and glucagon-stained slides (i.e. 6 slides by patient, 3 from the head and 3 from the tail).On each slide, three different areas were evaluated: the total pancreatic tissue area, the non-exocrine-non-endocrine pancreatic tissue area and the Fast Red-stained area, the latter corresponding either to alpha-cells or to beta-cells area, depending on the immunostaining.In R software, a picture is converted into 3 matrices, corresponding respectively to the blue, green and red color levels, each matrix element accounting for one pixel. The principles used for the selection of the pixels, belonging to the 3 predefined areas were the following.For total pancreatic tissue area delineation, the first step consisted in resizing the original picture to one hundredth of its full-resolution initial size. In other words, the original picture, whose size was between 1x108 and 6x108 pixels, was divided into squares of 10-pixel sides, each of these squares being integrated, in the resized picture, into one single central pixel, whose blue, green and red color levels were respectively the mean of blue, green and red levels from the 10x10 corresponding pixels in the original picture. This step allowed the smoothing of the existing discontinuities in the original picture to define a continuous pancreatic tissue surface (S2 Fig). The next step consisted then in the selection of any colored pixel, i.e. any pixel whose blue, green and red color levels were superior to the respective color levels of background pixels in the resized picture (Fig 1). The main difficulty in automating this process was to be able to take into account the differences in contrast and staining intensity between slides. The choice was therefore made to build an R script aimed at generating several propositions for total pancreatic tissue selection, by varying the color level threshold used by the different filters applied in the script (S1 File). An additional step of visual selection of the most accurate proposition by one investigator (FBS), in comparison to the original picture, was further carried out, allowing a critical step of visual control in the automated process (Fig 1).Open in a separate windowFig 1Example on the slide 14262.The first R script, used for the selection of the pixels belonging to total pancreatic tissue, generated 22 different propositions. The choice of the investigator is framed in red and the subsequent result indicated under the picture. An example of background pixel has been also pointed out on the picture miniature.For exocrine+endocrine pancreatic tissue area delineation, the first step also consisted in resizing the original picture, like in the total area delineation, but to a bigger size than for total pancreatic tissue, i.e. to one twentieth of its initial size. The next step consisted in the selection of either hematoxylin-stained blue pixels, i.e. colored pixels, whose blue level was superior to green and red levels, or to diaminobenzidine, brown, pixels or Fast Red, red, pixels, defined as colored pixels, whose red level was superior to green level. The R script intentionally generated automatically several propositions for exocrine+endocrine pancreatic tissue selection (S2 File). The choice of the most adequate selection by the investigator added a second step of visual control in the automated process (Fig 2).Open in a separate windowFig 2Example on the slide 14262.The second R script, used for the selection of the pixels belonging to functional pancreatic tissue, generated 16 different propositions. The choice of the investigator is framed in red and the subsequent results indicated under the picture.Given the small size of Fast Red-stained regions, the use of the full resolution picture was necessary for appropriate selection of the corresponding area. However, given the large size of the original picture, this required its split into several smaller pictures to allow easy manipulations within R software. The final Fast Red-stained selected area was then calculated by summing the results obtained on each small picture. The selection principle of Fast Red-stained pixels relied on the selection of colored pixels whose red level was higher than blue and green levels. However, the main difficulty encountered in this process was the presence of light pink artifact pixels on some slides, most often organized into large very slightly colored sheets, but sometimes with a color intensity similar to that of some weakly Fast Red-stained alpha- or beta islet cells. Thus, to gain detection sensitivity without losing specificity, a multi-step selection R-script was built (Fig 3) (S3 File). Briefly, assumption was made that every truly positive Fast Red-stained pixel should either display a high red level itself or a high red level pixel in its vicinity; artifact pixels being low in intensity and located in expanded regions only containing low red level pixels. Hence, the first step of the script selected high red level pixels (Fig 3B); the second step drew circles with a 50-pixel radius around each selected pixel (Fig 3C), thus pinpointing the regions where true Fast Red-stained pixels were located; and the final step identified the red pixels with a lower threshold within these circles as Fast Red-stained pixels (Fig 3D). The main drawback of this approach should be the inappropriate selection of islet neighboring tissue submitted to bleeding of the overstaining existing on some slides. However, the validation of the present methodology on some well-established parameters, as beta-cell mass, suggested that this effect was negligible (Fig 4). The whole process of picture analysis has been summarized on a flowchart (S3 Fig) (S4 File).