Multi-Lesion PET Radiomic Phenotyping and Aggregation for Prognostication In Diffuse Large B-cell Lymphoma Treated with CAR-T Therapy
Abstract
Purpose
To develop and evaluate a radiomic phenotyping and multi-lesion aggregation scheme to derive patient-level biomarkers for identifying progressive disease and predicting time-to-progression (TTP) in diffuse large B-cell lymphoma (DLBCL).
Methods
Lesions (n=1579) from 76 DLBCL patients treated with chimeric antigen receptor T-cell therapy (25 non-progressive) were segmented on pre-treatment FDG-PET/CT using PET-edge and PET-edge+ (MIM Software). SUV metrics (max/mean/min), volume, shape, and first-order radiomic features were extracted. Texture features were computed only for macro lesions (≥3 voxels per dimension). Scanner-related batches were harmonized using ComBat (log-ComBat for SUV, volume and skewed texture features; standard ComBat otherwise). Texture harmonization was applied to macro lesions, and micro-lesion textures were imputed using macro medians. After correlation filtering (Spearman ρ>0.8), radiomic phenotypes were derived via repeated K-means (k=2–15) followed by agglomerative clustering (consensus threshold=0.5). Thirteen phenotypes were identified; those in <25% of patients were grouped as “Rare,” yielding 7 stable phenotypes. Lesions were grouped by anatomical site (liver,bone, lymph node, soft tissue) and phenotype to generate patient-level features, including lesion counts and aggregated SUV/volume summaries (min/max/mean and Shannon entropy) at whole-body, organ-, and phenotype-specific levels. Clinical-only, aggregated-only, and combined models were compared for progression classification and TTP prediction using nested stratified cross-validation with SMOTE, block-wise feature selection (8 total), grid search, and 1000 bootstraps.
Results
Combined models using clinical+aggregated features (without prevalence-consolidation) performed best in prediction of both progression and TTP. For progression prediction, top model used SFS and MRMR feature selections for clinical and aggregated sets with SVM, achieving Matthews-correlation-coefficient of 0.38 (ROC-AUC=0.71). For TTP, highest C-index (0.67) was achieved by MI feature selection for both feature sets with GLMB.
Conclusion
Radiomic phenotyping with hierarchical aggregation across all lesions provides patient-level biomarkers that improve DLBCL prognostic modeling beyond clinical variables, with aggregated features adding predictive value for predicting both progression and TTP.