Radiomics-Based Machine Learning for Multi-Modality Tumor Classification and Prognosis in Lymphoma

Abstract

Purpose: This study aimed to develop a radiomics-based machine learning framework capable of differentiating tumor, normal, and mixed tumor-normal regions in lymphoma patients using 18F-FDG PET/CT images and to evaluate its effectiveness in predicting prognosis, including recurrence and mortality. Materials and Methods: F-18 FDG PET/CT imaging data from 60 patients diagnosed with lymphoma were retrospectively analyzed. A total of 417 radiomic features were extracted from each imaging modality (PET and CT) based on manually delineated tumor (n = 800) and normal tissue (n = 4,150) volumes of interest. Five machine learning classifiers—AdaBoost, Decision Tree, Gradient Boosting, Random Forest, and XGBoost—were trained using four distinct feature sets: PET radiomics features alone, CT radiomics features alone, combined PET/CT radiomics features, and standardized uptake value (SUV)-based metrics derived from PET images. To enhance tumor characterization, a scoring system integrating ensemble model predictions with anomaly detection using the Isolation Forest algorithm was developed. For prognostic modeling of five-year recurrence and overall survival, SUV-derived metrics, clinical variables, and Synthetic Minority Over-sampling Technique (SMOTE) were utilized to address class imbalance. Model generalizability and robustness were evaluated via external validation using an independent cohort consisting of 16 patients. Results: The CT-only radiomics model achieved the highest tumor classification performance with an AUC of 0.9690, compared to combined PET/CT radiomics model (AUC: 0.9639) and PET radiomics model (AUC: 0.9607), while PET-only radiomics model demonstrated optimal sensitivity (recall: 65.80%). XGBoost consistently outperformed other algorithms across all feature combinations, with PET/CT achieving 94.23% accuracy and PET-only achieving 93.47% accuracy. For prognostic prediction without clinical data, recurrence accuracy ranged from 42-67% (without SMOTE) to 50-75% (with SMOTE), while mortality prediction ranged from 71-79% (without SMOTE) to 71-86% (with SMOTE). However, clinical data integration yielded inconsistent results, with recurrence prediction accuracy ranging from 47% to 92%. External validation confirmed model generalizability, with PET-based features showing the best performance (accuracy: 90.34%, AUC: 0.8852). Sensitivity decreased from 65.80% to 40.4% in external validation, indicating inter-institutional variability and the need for institutional calibration. Conclusion: The developed radiomics-based machine learning framework effectively differentiates tumor, normal, and mixed volumes in lymphoma patients, demonstrating strong potential for enhancing prognosis prediction. However, sensitivity reduction in external validation underscores the need for further refinement and institutional calibration before widespread clinical adoption.