INTRODUCTION: Until now, diagnosis the subtype of pancreas cystic lesion remains a major challenge. The accuracy of preoperative invasive diagnosis procedures is still very difficult due to the low cellularity of the aspirate. Accuracy verification of nature lesion’s is essential for predicting tactical planning and planning interventions.OBJECTIVE: To determine the diagnostic significance of texture analysis of contrast enhancement СT in differentiation of tumor and tumor-like pancreatic cystic lesions; to compare results of application of two models (2D and 3D) segmentation of CT images; to develop a diagnostic model including texture features to differentiate tumor and tumor-like pancreatic cystic lesions.MATERIALS AND METHODS: Clinical and CT data of 40 patients with pancreatic cystic lesions were collected for this study. Among these patients, 15 were pathologically diagnosed with serous cystadenoma, 15 were diagnosed with mucinous cystadenoma and 10 were diagnosed with pseudocyst. The radiomic features were extracted from four CT phases (native, arterial, venous and delayed). All images were normalized prior to the radiomics analysis, using spatial resampling with fixed voxel size of 1 mm3 (RES) and density threshold from 0 to 200 HU. For each phase, one radiologist (3 year`s experience in abdominal imaging) segmented the lesion contour on each slice (3D) and on the slice with maximum axial diameter (2D).Statistics: The program R 4.3.2 (R Foundation for Statistical Computing, Vienna, Austria) was used. The Mann-Whitney test and AUC with 95% confidence interval were used to assess the discriminative texture predictors for tumour types. Texture features were included in the analysis after standardization, and L1 (LASSO) regularization was used to select predictors. Finally, discriminative models were evaluated by bootstrap estimation and Matthews correlation coefficient.RESULTS: Using 3D segmentation is preferable for differentiation of tumor and tumor-like pancreatic cystic lesions. A 2-D radiomics diagnostic model was included features (INTENSITY-HISTOGRAM_IntensityHistogram75th Percentile, MORPHOLOGICAL_Volume, INTENSITY-BASED_StandardDeviation) from native and arterial phases. It was resulted in an average AUC 0.89, with an sensitivity and specificity 70 and 93.3% according to pseudocysts, 73.3 and 92% according to mucinous cystadenomas, 86.7 and 80% according to serous cystadenomas. A 3-D radiomics diagnostic model was included features (MORPHOLOGICAL_SurfaceToVolumeRatio, INTENSITY-BASED_StandardDeviation, GLCM_Correlation, GLSZM_ZonePercentage) from native, arterial and delayed phases. It was resulted in an average AUC 0.96, with an sensitivity and specificity 80 and 96.7% according to pseudocysts, 86.7 and 88% according to mucinous cystadenomas, 80 and 88% according to serous cystadenomas. DISCUSSION: Currently, textural analysis is aimed at solve two main problems — differentiation of histological classes and grade of pancreatic cysts. The standardization of pre-processing and segmentation remains an unresolved issue. At the time of this study, we haven`t found any papers analyzing all the phases of CT imaging. A review of publications revealed that in the majority of cases researchers analyzed only one phase (arterial/venous) by 3D-segmentation. In our study, four phases of CT (native, arterial, venous and delayed) were analyzed by two types of segmentaion. In order to reduce texture ranges and offset the segmentation errors, we investigate preprocessing steps such as density distribitions (0–200 HU) and voxel resampling 1 mm3 (RES). In contrast to other papers, in our study there are no statistically significant textural features for the venous phase. Also, we don`t identify higher-order textural features as a differentiation predictors.CONCLUSION: Texture analysis of contrast enhancement СT have a favorable differential diagnostic performance for tumor and tumor-like cystic lesions of the pancreas.
