Identification of pancreatic nonfunctional neuroendocrine tumors and solid pseudopapillary tumors via the construction of a consensus clustering model based on enhanced CT images

Background : Unsupervised clustering has played a greater role in the diagnosis and differential diagnosis of pancreatic tumors in recent years. This study aimed to investigate the value of constructing a c[1]lustering model for unsupervised learning based on enhanced CT to identify pancreatic nonfunctional neuroendocrine tumors (NF-pNETs) and solid pseudopapillary tumors (SPTs). Methods : 45 patients with SPTs and 47 patients with NF-psNETs were retrospectively analyzed. The data were randomly divided into a training set and a validation set at a ratio of 7:3. One-way logistic regression was performed for each clinical variable to assess its relationship with the clustered labels, and a logistic regression model was fitted with the clustered labels as the dependent variable and the clinical variables as independent variables. Variables with P values <0.1 were selected for further analysis. Multifactorial logistic regression models were fitted via the clinical variables selected in the univariate analysis, ridge regularization (L2 penalty) was used to prevent overfitting and address potential multicollinearity, with the strength of regularization (alpha) set to a default value of 1. Clinical variables that were meaningful in the multifactorial logistic regression analyses were used to construct the final logistic regression model, which was used to predict individual clinical-based group labeling on the basis of individual clinical characteristics. For imaging, the optimal number of clusters k selected by the covariance coefficient was used to build the clustering model via unsupervised classification of the lesions through consensus clustering analysis fusing the imaging histological features of arterial-phase, venous-phase, and delayed-phase images. The clinical factors with a final p value < 0.2 were subsequently combined with the clustering model to perform stepwise multivariate logistic regression analyses, thereby establishing a joint model. The performance of the three models was assessed via AUC values, and column line plots were generated to visualize the models. Finally, the clinical validity of the models was assessed via decision curve analysis (DCA). Results : The AUCs of the clustered and clinical models were 0.70 and 0.74 (95% CI: 0.66–0.81) in the training set and 0.65 and 0.75 (95% CI: 0.64–0.87) in the validation set, and the AUCs of the training and test sets in the joint model were 0.88 (95% CI: 0.81–0.94) and 0.85 (95% CI: 0.71–0.96), respectively. Decision curve analysis revealed that the joint clinical-clustering model had a greater net benefit when the high-risk threshold probability was in the range of 0--1. Conclusions : Unsupervised clustering models based on enhanced CT have potential for discriminating between SPTs and NF-PNETs, which can inform clinical decision-making.