Malignancy Predicting Model of Pancreatic Cystic Neoplasm in Contrast-Enhanced CT by Machine-Learning and Radiomics

초록

Objectives: Accurate identification of PCNs is crucial due to their diverse characteristics. To improve diagnostic precision, this study employs radiomics and machine learning to develop a predictive model for PCN malignancy risk, utilizing CT imaging data. Materials and Methods: This retrospective study included 71 pts who underwent op for PCN and had CT scans within 90 days before op. PCNs were categorized as benign or malignant based on pathological results. We conducted texture analysis on segmented cystic fluid using an image processing program, calculating the standard deviation of gray-scale histograms(SDGSH) within the segmented area to detect increased heterogeneity. From the volume of interest in PCNs, 824 radiomics features were extracted. Radiomics and clinical features were selected using a random forest regressor, and the PCN classification model was trained and validated using a random forest classifier with 20-fold stratified cross-validation. Results: The SDGSH representing texture heterogeneity, increased progressively from the non-dysplasia group to the invasive carcinoma group, with significant differences noted. The SDGSH demonstrated good predictive power for malignancy in pre-op CT images, with an AUC of 0.81 (95% CI 0.67-0.95). An optimal cut-off value of 17.597 yielded 87.5% sensitivity and 69.6% specificity. In internal validation, the radiomics-based machine learning model showed a meaningful AUC of 0.83 (95% CI 0.72-0.94), with a sensitivity of 81.25% and a specificity of 64.10%. Conclusions: Radiomics technology, specifically SDGSH correlating with textural heterogeneity, and machine learning models analyzing cyst texture heterogeneity in CT images, were found to be good predictors of malignancy risk in PCN.