Pancreatic neuroendocrine tumors (PNETs) originate from neuroendocrine cells of the pancreas, representing about 3% of all pancreatic tumors. The WHO classifies PNETs into grade 1 (well-differentiated), grade 2 (intermediately differentiated), and grade 3 (poorly differentiated) based on mitotic rate and Ki-67 proliferation index. Tumor grading is a critical prognostic indicator, as G1 tumors are typically managed with active monitoring while G2/3 tumors require more aggressive surgical treatment.
The current gold standard for pathological grading relies on postoperative pathological examination of surgical specimens. Preoperative grading through endoscopic ultrasonography-guided fine-needle aspiration or biopsy (EUS-FNA/B) is hindered by invasiveness, limited accuracy, difficulty capturing tumor heterogeneity, and a high technical threshold. There is therefore a pressing need for precise, non-invasive techniques to classify G1 and G2/3 PNETs before surgery.
Endoscopic ultrasonography (EUS) is commonly used to diagnose PNETs and is considered a highly accurate imaging tool due to its ability to produce detailed images of pancreatic lesions. According to European Neuroendocrine Tumor Society (ENETS) guidelines, EUS is the preferred imaging method when other tests are inconclusive, demonstrating superior efficacy compared to CT, MRI, and abdominal ultrasonography for detecting PNETs.
Ultrasomics is ultrasound-based radiomics that extracts quantitative features from medical images. Previous studies have shown that radiomics techniques can effectively predict grading of PNETs on CT and MRI scans. Additionally, features from ultrasomics analysis of B-mode ultrasound images show potential in predicting pathological grading, and peritumoral radiomics features have demonstrated significant associations with tumor-related outcomes.
This retrospective study included 81 patients (49 women, 32 men) with PNETs confirmed through pathological examination, comprising 51 with grade 1 and 30 with grade 2/3 tumors. Patients were enrolled from October 2013 to January 2024 at the First Affiliated Hospital of Guangxi Medical University. They were randomly allocated to training (n=48) or test (n=33) groups in a 6:4 ratio.
Inclusion criteria required preoperative EUS scan confirmation of PNETs with pathology-confirmed grading, clear EUS images before biopsies, and no prior chemotherapy or radiotherapy. Seven patients were excluded due to inability to display the entire lesion, significant motion artifacts or noise, or combined other tumor types. EUS images were acquired using a standard dynamic procedure with the EU-ME2 device (Olympus) and SU-9000 device (FUJIFILM).
Two EUS experts manually delineated the intratumoral region of interest (ROI) using ITK-SNAP software on DICOM images converted to nii.gz format. The peritumoral ROI was obtained by expanding the intratumoral ROI by 3 mm using morphological dilation. Three separate ROI images were chosen per patient: intratumoral, peritumoral, and combined ROIs. Features were extracted using PyRadiomics across seven categories totaling 107 ultrasomics features.
Feature selection involved multiple steps: a Mann-Whitney U test filtered features at p<0.05, Spearman rank correlation removed redundant features with coefficients exceeding 0.9, and a greedy recursive deletion method further refined the feature set. The LASSO regression model with 10-fold cross-validation identified features with nonzero coefficients. This process retained 4 intratumoral, 6 peritumoral, and 6 combined ultrasomics features for model construction.
A Multilayer Perceptron (MLP) machine learning algorithm was employed to create classification models for differentiating G1 and G2/3 PNETs. After LASSO feature selection, the retained intratumoral, peritumoral, and combined ultrasomics features were used to build three separate ultrasomics models. A clinical model was also constructed using shape and maximum diameter, identified through multivariate logistic regression analysis.
Multiple machine learning algorithms were evaluated including Random Forest, XGBoost, ExtraTrees, Support Vector Machine, LightGBM, Logistic Regression, and K-Nearest Neighbors. The MLP model demonstrated superior performance with greater consistency between training and test groups (training AUC=0.833, test AUC=0.800 for intratumoral), while tree-based models like Random Forest, XGBoost, and ExtraTrees exhibited overfitting tendencies. All models were established using 5-fold cross-validation with hyperparameter tuning.
Among all models tested, the combined ultrasomics model integrating both intratumoral and peritumoral features achieved the greatest performance. In the training group, it reached an AUC of 0.858 (95% CI 0.7512-0.9642) with accuracy of 0.729, sensitivity of 0.765, and specificity of 0.710. In the test group, it achieved an AUC of 0.842 (95% CI 0.7061-0.9785) with accuracy of 0.758 and sensitivity of 0.800.
The intratumoral model achieved an AUC of 0.833 in training and 0.800 in testing, performing comparably to the clinical model (training AUC=0.810, test AUC=0.762). The peritumoral model showed AUC values of 0.787 (training) and 0.788 (test). Notably, the combined model outperformed the clinical model in the test group based on the DeLong test, confirming the added value of integrating both tumor regions.
The calibration curves for all ultrasomics and clinical models demonstrated high consistency between predicted probability and actual G2/3 PNET outcomes in both training and test groups. The Hosmer-Lemeshow test further validated the models with p-values of 0.424 (training) and 0.209 (test) for the combined model, confirming reliable calibration across all models.
Decision curve analysis (DCA) showed that the combined ultrasomics model provided significant improvement in patient intervention efficacy compared to treat-all or treat-none strategies in both cohorts. The net benefit of the combined model appeared to outperform both the clinical model and other ultrasomics models. SHAP values were used to visualize feature importance, with density plots illustrating how individual feature changes influenced G1 versus G2/3 classification predictions.
Most existing radiomics literature on PNETs focuses exclusively on intratumoral regions while neglecting the peritumoral area. However, previous studies have demonstrated significant predictive capabilities of peritumoral radiomics models for pathological outcomes, lymph node metastasis, and recurrence risk stratification. This study found that combining peritumoral and intratumoral features yielded synergistic effects in distinguishing G1 and G2/3 PNETs.
G2/3 PNETs demonstrate notably heightened aggressiveness compared to G1 tumors, with increased susceptibility to lymph node and microvascular metastasis. The variances in micro invasion and immune cell infiltration surrounding the tumor could potentially explain the disparities in ultrasomics characteristics of EUS in the peritumoral vicinity. These microenvironment differences are captured by the peritumoral features, providing valuable prognostic insights that complement the intratumoral analysis.
Accurate preoperative determination of PNET grading is crucial for clinical treatment planning. G1 asymptomatic and non-functional PNETs under 2 cm are recommended for active monitoring, while G2/3 tumors are categorized as high-risk and typically necessitate more aggressive surgical treatment. The combined ultrasomics model offers a non-invasive approach to predict pathological grades before surgery, potentially reducing the need for invasive EUS-FNA/B procedures.
The EUS-based combined ultrasomics model with SHAP interpretability offers significant predictive value before fine-needle aspiration and surgery. By integrating radiomics and machine learning with routine EUS imaging, clinicians can obtain objective, quantitative grading predictions that overcome the subjective interpretation limitations of conventional macroscopic anatomical imaging features. This approach supports personalized treatment planning by identifying high-risk patients who would benefit from aggressive intervention.
The integrated model combining EUS ultrasomics features from both intratumoral and peritumoral tumor regions accurately predicts PNET pathological grades before surgery, aiding personalized treatment planning. The combined MLP-based model achieved the best overall performance with an AUC of 0.858 in training and 0.842 in testing, outperforming models using individual regions or clinical features alone.
The study has several limitations including its single-center retrospective design with a relatively small sample size of 81 patients, which may limit the generalizability of results. The manual segmentation of ROIs by specialists, though validated with high inter-observer consistency (ICC>0.8), introduces potential variability. Future research should incorporate multicenter prospective studies with larger sample sizes and explore automated segmentation approaches to enhance clinical applicability.