Stacking Ensemble Learning-Based [¹⁸F]FDG PET Radiomics for Outcome Prediction in Diffuse Large B-Cell Lymphoma

Area of Science:

• Oncology
• Medical Imaging
• Machine Learning

Background:

• Diffuse large B-cell lymphoma (DLBCL) is an aggressive non-Hodgkin lymphoma requiring accurate prognostic models.
• Current prognostic tools may not fully capture the heterogeneity of DLBCL.
• Integrating advanced imaging features with clinical data can potentially enhance outcome prediction.

Purpose of the Study:

• To develop and validate an analytic approach using [18F]FDG PET radiomics and stacking ensemble learning for improved outcome prediction in DLBCL.
• To compare the performance of radiomics-only and combined clinical-radiomics models.
• To assess the model's ability to stratify patient risk more effectively than the International Prognostic Index.

Main Methods:

• Radiomics features were extracted from pretreatment [18F]FDG PET scans of 240 DLBCL patients using four semiautomatic segmentation methods.
• Features were harmonized, and the most reliable ones were selected using correlation analysis and the LASSO algorithm.
• A stacking ensemble learning approach combined machine learning classifiers to build predictive models for progression-free and overall survival.

Main Results:

• The SUV4.0 segmentation method demonstrated the highest interobserver reliability, yielding 830 radiomics features.
• The combined clinical-radiomics model built with stacking ensemble learning achieved the best discrimination performance.
• In the external testing set, the combined model showed an AUC of 0.771 for progression-free survival and 0.725 for overall survival, outperforming the International Prognostic Index.

Conclusions:

• A combined model integrating [18F]FDG PET radiomics and clinical data using stacking ensemble learning offers improved risk stratification for DLBCL patients.
• This approach enhances prognostic accuracy beyond traditional methods.
• The findings suggest potential for more personalized treatment strategies in DLBCL.