PET Intra-Patient Inter-Lesion Radiomics Feature Aggregation to Enhance PRRT Progression Predictions in Patients with Neuroendocrine Tumors

Maziar Sabouri; Fereshteh Yousefirizi; Omid Gharibi; Ayca Dundar; Camila Zamboni; Sanchay Jain; Yusuf Menda; Arman Rahmim; Ahmad Shariftabrizi

Introduction: Peptide Receptor Radionuclide Therapy (PRRT) marks a significant advance in treating advanced neuroendocrine tumors (NETs), yet its efficacy remains difficult to predict. The FDA-approved [177Lu]Lu-DOTA-TATE targets somatostatin receptor-positive gastroenteropancreatic NETs that have progressed on somatostatin analogs. Pre-treatment PET imaging often reveals multiple lesions and may hold untapped predictive value for optimizing individualized therapy. Radiomics features extracted from tumors in somatostatin receptor PET/CT images may offer predictive insights into the therapeutic response of NET patients undergoing [177Lu]Lu-DOTA-TATE PRRT. Focusing solely on the largest or most active lesions may lead to underperforming models in survival studies. In this study, we evaluated different aggregation approaches for radiomic features, aiming to enhance PRRT progression prediction (PP). Methods: In this study, 74 NET patients (18 progression-free) who underwent [177Lu]Lu-DOTA-TATE therapy were retrospectively analyzed. On the pre-treatment SSTR PET/CT scans, all metastatic lesions were manually segmented using PET Edge workflow (MIM Software, Inc.), and standardized uptake values (SUV) were normalized between 0 to 20. Subsequently, lesions were sorted based on intensity and volume. Intensity sorting was performed twice using SUVmax and SUVmean. This process resulted in three sorted lists from which the top 1, 3, and 5 lesions were selected. Radiomics including shape, intensity, and texture were then extracted separately from each lesion per patient using the Pyradiomics library and aggregated globally as well as across the selected lesions using statistical measures including mean, median, min, max, variance, skewness, and kurtosis. Additionally, for the top 3 and top 5 lesions, features from the selected lesions were stacked. The data were then normalized and went through a nested cross-validation with 3-fold outer and inner loops. The class imbalance was addressed by SMOTE. Highly correlated features were removed based on Pearson correlation (threshold = 0.95), and the 10 most predictive feature combinations were selected iteratively using the Minimum Redundancy Maximum Relevance (MRMR) algorithm. In each fold, the selected features were fed to 8 machine learning (ML) algorithms: Random Forest (RF), Naive Bayes (NB), Decision Tree (DT), eXtreme Gradient Boosting (XGB), Multilayer Perceptron (MLP), Logistic Regression (LR), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN) to perform a binary classification (progression vs. non-progression). Models were tuned using grid search and evaluated using area under the characteristic curve (AUCC), accuracy (ACC), precision, recall, F1-score, balanced accuracy, and specificity. The performance of the models was compared using the Kruskal-Wallis test. Results: Table 1 sorts the 10 models with the highest performance in PP. It is evident that models using aggregated radiomics data generally performed better than those based on a single lesion. Feeding DT with the min value of features extracted from the top three lesions sorted by volume (Min-Volume-Top3) led to the highest performance (AUCC = 0.75, ACC = 0.82). KNN trained on the not-aggregated features of the top lesion, sorted using SUVmean (SUVmean-Top1), was the only single-lesion model among the top 10 models. Despite better performance gained by feature aggregation, a Kruskal-Wallis test revealed no statistically significant differences between any of the models in the table. Comparisons among the aggregation approaches (after Bonferroni correction) did not indicate statistically significant differences in either AUCC or ACC across model pairs. Conclusions: This study demonstrates that lesion aggregation tends to improve performance in predicting disease progression for NET patients undergoing PRRT. The DT classifier fed with aggregated lesion data showed the best results. These findings highlight the need for further research to refine such approaches.

PET Intra-Patient Inter-Lesion Radiomics Feature Aggregation to Enhance PRRT Progression Predictions in Patients with Neuroendocrine Tumors

View Online

Abstract

Details

Metrics