In all cases of segmentation the following was observed: the threshold of 0.9 (excellent reliability) for the arterial phase of the study was not achieved when the border of the pancreas was hyposegmented by more than 2 mm and hypersegmented by more than 1 mm; for the portal phase of the study the threshold is 0.9 was not achieved when the border of the pancreas was hyposegmented by more than 3 mm and hypersegmented by more than 2 mm. Significant statistical differences were found between ICC corresponding thresholds for the arterial and venous phases of the study (p<0,001).

The results are presented in tables 1, 2.

Table 1: The table shows the values ​​of the ICC for each offset from the boundary of pancreas.

Table 2: The table shows the variability ​​of the ICC values for each offset from the boundary of pancreas. Statistical differences between ICC values of 0.1 from the threshold were found to have a strong negative impact on segmentation reproducibility.

Thus, the texture analysis technique is more sensitive to include areas outside the desired zone of interest. This is explained, first of all, by the fact that beyond the boundaries of these structures, more heterogeneous tissue is most often located relative to the more homogeneous structure of the pathology. In our example, the expansion of the area of interest was associated with the inclusion of peripancreatic tissue and vessels, which expand the range of encountered pixels from negative HU values to highly bright ones in the vessels, especially in the arterial phase. From this we can conclude that hyposegmentation is preferable, since it preserves the pattern “recognizable” by texture analysis without the effect of “averaging”, which is inevitable with hypersegmentation. But any area of under-marking is a loss of valuable diagnostic information. Moreover, the structure of the peripheral parts of the zones of interest may be the most informative for solving a number of problems of texture analysis, since it may include unique and most characteristic patterns. To avoid this, methods for analyzing peritumoral tissue and habitat radiomics are being actively developed [10, 11, 12]. Considering everything that has been said, the approach of preference for hyposegmentation is appropriate only in the case of a correct preliminary assessment of the size and location of near-by tissues to the area of lesion.