Segmentation is an important process for AI radiomics, radiotherapy planning, assessing the dynamics of the pathological process, and radiomics. Texture analysis is a technique with high potential for use in clinical diagnostic practice. Texture analysis uses images obtained during routine diagnostic procedures, but involves a group of mathematical calculations performed on the information contained within the images [1]. This method is especially useful in the analysis of CT studies of the pancreas, since classical medical imaging methods face some significant difficulties. For example, computed tomography has low sensitivity for detecting and differentiating small volume pancreatic lesions. In addition, the heterogeneity of the many types of pancreatic lesions leads to difficulties in determining malignancy. As a result, clinicians are forced to use invasive diagnostic methods, which can also give uncertain outcomes [2]. In such circumstances, thanks to its non-invasiveness and high accuracy, the texture analysis method could be useful.

But many questions about the reproducibility of the obtained radiomic features still remain. Reproducibility refers to the consistency between the results of repeated measurements of data (in this case, segmentation results) using the same methodology [3].

Researchers cite the tendency for high variability in manual segmentation as one of the reasons for low reproducibility [4,5]. In medical image segmentation, it is well known that inter- and intra-operator variability exists. The first refers to the observed differences in segmentation results obtained by two different operators, while the second refers to the observed differences between two results of segmentation tasks performed by the same operator at different times [3]. Actuality of this problem is especially high for complex organs such as the pancreas [6].

Conduction of radiomic research is complicated by the complexity of the segmentation. The main difficulties are the time consumption and the requirements for not only radiological skills, but also skills in working with segmentation software. Even so, serious questions remain regarding the search for a standard, because correct segmentation is influenced by many factors not related to the radiology experience of the expert. For example, experience with segmentation tools, burnout, operators’ well-being, and various other reasons are no less significant factors affecting the quality of the region/volume of interests.

Automated and semi-automated segmentation tools are promising instruments for texture analysis of medical images of the pancreas. But achieving accurate segmentation with these methods is also a difficult task due to the structural diversity of the gland, differences in its location in the abdominal cavity, close proximity to other organs such as the duodenum and gallbladder [7,8].

Ultimately, choosing the optimal segmentation is a search for a compromise. In artificial intelligence tasks, this is a difficult process, because everything depends on a specific task, and there are many scenarios for its use. In turn, in the tasks of texture analysis there is an obvious indicator for compromise - the reproducibility of texture features depending on errors in the segmentation. In this study we tried to evaluate the impact of said errors on segmentation variability.