Primary Mediastinal Large B-Cell Lymphoma (PMBCL) is a rare and distinct subtype of aggressive B-cell lymphoma with a global incidence of approximately 0.4 per million. It typically arises from thymic B-cells, predominantly affects young adults, and presents as a large anterior mediastinal mass. Despite generally favorable outcomes following rituximab-based chemoimmunotherapy (R-CHOP), roughly 5-10% of patients develop refractory disease associated with poor prognosis, and most relapses occur within two years of initial treatment, sometimes involving extranodal sites.
Current prognostic tools and their limits: End-of-treatment (EoT) PET/CT assessment using the five-point Deauville Score (DS) is the standard method for gauging treatment response in PMBCL and other lymphomas, as endorsed by the International Conference on Malignant Lymphomas Working Group. However, the DS is a subjective visual score, and ambiguity in defining DS = 4 versus DS = 5 creates interobserver variability in the clinically critical zone. Single-time-point assessment also ignores the trajectory of metabolic change across treatment, which may carry additional prognostic information.
What radiomics adds: Radiomics extends traditional imaging analysis by extracting hundreds of quantitative features from segmented tumor volumes on PET and CT, including shape descriptors, first-order intensity statistics, and higher-order texture features capturing intratumoral heterogeneity. Delta radiomics specifically computes the change in these features between two scan time-points (in this case, pre-treatment baseline and EoT), capturing the tumor's response trajectory rather than its state at a single moment. Prior studies in non-small cell lung cancer, gastric cancer, esophageal cancer, and head-and-neck cancers have demonstrated that delta radiomics improves outcome prediction over single-time-point radiomics, and this study is among the first to systematically evaluate this approach in PMBCL.
The authors from BC Cancer Research Institute (Vancouver) retrospectively analyzed 103 PMBCL patients treated with R-CHOP between 2005 and 2021. Because staging PET scans were not standard until 2011 due to resource constraints, only 31 patients had both pre-treatment and EoT scans suitable for delta radiomics analysis. The full EoT-only cohort extended to 50 patients (31 plus an additional 19 EoT-only cases) for single time-point analysis.
The study enrolled 103 PMBCL patients diagnosed between 2005 and 2021 at BC Cancer, all confirmed by expert central pathology review per WHO classification criteria and treated with curative-intent R-CHOP chemoimmunotherapy. PET/CT scans were acquired on two GE Healthcare scanners (Discovery 600 and Discovery 690) following a standardized protocol: 6-hour fast, blood glucose below 200 ng/dL, 300-400 MBq [18F]FDG injection, and a 60-minute uptake phase.
Dataset split: Of the 103 patients, only 31 had both baseline (pre-treatment) and EoT PET/CT scans, due to the absence of routine staging PET prior to 2011. These 31 patients formed the delta radiomics cohort. For single time-point EoT analysis, an additional 19 EoT scans were included, yielding 50 total EoT cases. The 50-case EoT cohort had a 24.0% progression rate and average follow-up of 5.56 years (standard deviation 3.74 years), while the 31-case paired cohort had a 15.6% progression rate and average follow-up of 3.73 years (SD 2.17 years).
Expert segmentation: Nuclear medicine physicians manually segmented primary tumor volumes on all 31 baseline and 50 EoT PET scans using the MIM software (version 7.2.3), blinded to patient outcome. Four segmentation methods were evaluated: 41% SUVmax thresholding, 25% SUVmax thresholding, manual contouring, and PET-Edge (gradient-based). Prior reproducibility studies by the same group established minimal interobserver variability and high intra-class correlation with the PET-Edge MIM method, which was adopted as the primary approach. Although PET and CT images are hardware-registered during acquisition on the hybrid scanner, an additional affine PET-CT registration step was performed to enable accurate CT segmentation.
Radiomic feature computation: Features were extracted using the Pyradiomics package. First-order features included SUVmax, SUVmean, metabolic tumor volume (MTV), total lesion glycolysis (TLG), and entropy of the mediastinal mass. Higher-order texture features included Neighbourhood Grey Tone Difference Matrix (NGTDM) descriptors (Busyness, Coarseness, Complexity, Contrast, Strength) and Grey Level Run Length Matrix (GLRLM) features (Grey Level Non-uniformity, Run Length Non-uniformity). The CSH (intensity-based area under the curve of cumulative SUV-volume histogram) was also computed. Grey level quantization used 64 bins for PET (0.3125 SUV units per bin) and 400 bins for CT (10 Hounsfield Units per bin). To capture spatial heterogeneity across tumor depth, Radial Intensity Mean (RIM) features were also extracted, representing average intensity per concentric voxel envelope from the tumor surface inward.
The study defined two variants of delta radiomics to capture treatment-induced change. Absolute delta is computed as the absolute value of the difference between EoT and baseline feature values: |EoT Radiomics - Baseline Radiomics|. Relative delta normalizes this difference by the baseline value: |EoT Radiomics - Baseline Radiomics| / Baseline Radiomics. Both variants were analyzed separately and in combination with baseline features, resulting in seven distinct feature set configurations: Baseline only, EoT only, Baseline + EoT, Relative Delta, Absolute Delta, Baseline + Relative Delta, and Baseline + Absolute Delta.
Scanner harmonization: Because 31 paired scans were acquired across two scanners (GE Discovery 600 and 690) at different time points, the authors first tested for scanner-induced additive effects (mean shifts, assessed by two-sample t-test) and multiplicative effects (heteroscedasticity, assessed by Bartlett's test). No statistically significant additive or multiplicative scanner effects were found, indicating harmonization was not strictly necessary. Nevertheless, Longitudinal ComBat was applied as a precautionary step, and Kolmogorov-Smirnov tests confirmed that harmonized and non-harmonized features showed no significant differences. The final analysis therefore used non-harmonized features.
Missing data imputation: Not-a-Number (NaN) feature values were imputed using mean imputation, balancing computational efficiency against accuracy for the relatively small dataset.
Machine learning workflow for progression prediction: Constant features and highly correlated features (Pearson correlation above 0.8) were removed. Feature selection used Sequential Feature Selection (SFS), a wrapper method that iteratively adds features that most improve model performance; this outperformed LASSO, RELIEF, and linear/kernel-based PCA in preliminary comparisons. Classifiers evaluated were K-Nearest Neighbor (KNN), Linear Discriminant Analysis (LDA), and Random Forest (RF), plus the ICARE (Individual Coefficient Approximation for Risk Estimation) method designed for small-sample settings. Class imbalance was addressed with SMOTE (Synthetic Minority Oversampling Technique). Model evaluation used nested cross-validation with 10 outer folds and 5 repetitions for outer CV, and 5 inner folds with 20 repetitions for inner hyperparameter tuning, applied to 85% of the data, with the remaining 15% held as an unseen test set.
Survival analysis for Time to Progression (TTP): CoxNet, which combines Cox Proportional Hazards regression with LASSO (L1) regularization, was used for TTP prediction. Feature selection for CoxNet evaluated SFS, Chi-Square, Mutual Information, Random Forest feature importance, and LASSO filter methods. Hyperparameter tuning for CoxNet used 10-fold cross-validation with GridSearchCV to optimize the L1 ratio, alpha_min_ratio, and max_iter parameters.
For single time-point EoT analysis on the 50-case cohort, LDA and RF performed best using PET and PET-CT features respectively. With PET-CT EoT features, RF achieved accuracy 0.92 plus/minus 0.02 and F1 score 0.91 plus/minus 0.02, while ICARE achieved accuracy 0.79 plus/minus 0.09 and F1 0.81 plus/minus 0.08. PET-CT features significantly outperformed PET-only features under RF (accuracy 0.87 plus/minus 0.03 for PET-only versus 0.92 plus/minus 0.02 for PET-CT, p less than 0.05), confirming that CT texture information adds discriminative value beyond metabolic PET data alone.
ICARE on paired scan feature sets (n=31): Using ICARE on PET-CT features across all feature combinations, Absolute Delta achieved accuracy 0.81 plus/minus 0.15 and F1 0.77 plus/minus 0.18, while Baseline + Relative Delta achieved accuracy 0.84 plus/minus 0.11 and F1 0.82 plus/minus 0.13. Both outperformed EoT-only features under ICARE (accuracy 0.78 plus/minus 0.14, F1 0.76 plus/minus 0.19), though most differences did not reach statistical significance given the small n.
ML classifiers on paired scan feature sets (n=31): Among the ML techniques evaluated on PET-CT features, the clearest performance gains from delta radiomics emerged. For Relative Delta features, RF achieved accuracy 0.89 plus/minus 0.04 and F1 0.87 plus/minus 0.05, while LDA achieved accuracy 0.89 plus/minus 0.03 and F1 0.89 plus/minus 0.03. For Absolute Delta, LDA achieved accuracy 0.89 plus/minus 0.03 and F1 0.89 plus/minus 0.03. The Baseline + Relative Delta combination under KNN achieved accuracy 0.86 plus/minus 0.03 and F1 0.87 plus/minus 0.03. These results represent a meaningful improvement over EoT-only ML results (RF accuracy 0.83 plus/minus 0.05 on the same n=31 subset), demonstrating the added prognostic signal from longitudinal change features.
Modality comparison: PET-CT consistently outperformed PET-only features across all feature sets and prediction methods, reinforcing that CT-derived texture and morphological features complement PET metabolic information. No statistically significant differences in segmentation method (41% SUVmax versus 25% SUVmax versus manual versus PET-Edge) were observed for any prediction task, confirming the robustness of the radiomic analysis to the choice of segmentation approach.
For time to progression (TTP) survival analysis, the concordance index (Harrell's c-index) was used as the primary metric. A c-index of 0.5 represents random prediction, while 1.0 is perfect concordance. ICARE showed limited efficacy for TTP prediction using baseline-only features (c-index below 0.4), but improved when delta and EoT features were incorporated.
Best TTP results on paired scans (n=31): Using Baseline + Absolute Delta PET-CT features, CoxNet with LASSO achieved c-index 0.67 plus/minus 0.06, and CoxNet with SFS achieved c-index 0.65 plus/minus 0.17. ICARE on the same feature set reached c-index 0.61 plus/minus 0.11. For EoT-only features on the same 31-case subset, CoxNet with LASSO achieved c-index 0.65 plus/minus 0.07, indicating that Baseline + Absolute Delta performed comparably to or slightly better than EoT-only features.
Expanded EoT cohort (n=50): On the larger 50-case EoT cohort, performance improved: CoxNet with SFS reached c-index 0.68 plus/minus 0.09 and CoxNet with LASSO reached 0.67 plus/minus 0.09, both statistically better than ICARE (c-index 0.65 plus/minus 0.23, p less than 0.05). The trend toward better TTP prediction with more EoT cases was present but did not achieve statistical significance when comparing the n=31 and n=50 EoT subsets directly.
Deauville Score comparison: The clinical Deauville Score (DS) achieved c-index 0.66 plus/minus 0.09 for n=50 and 0.67 plus/minus 0.03 for n=31. Critically, no statistically significant difference was detected between the radiomics-based TTP models and the DS (p greater than 0.05 across comparisons). This finding suggests that within the constraints of this dataset, delta radiomics and EoT radiomics provide prognostic information comparable to the current clinical standard, without demonstrating clear superiority. The authors caution that the small sample sizes limit the power to detect moderate effect sizes.
Beyond binary relapse/progression classification and survival analysis, the study explored a novel regression task: predicting the Total Metabolic Tumor Volume (TMTV) on EoT scans from baseline PET-CT radiomic features. This question asks whether the metabolic burden remaining after treatment can be estimated from pre-treatment imaging characteristics, potentially offering an early indicator of treatment inadequacy before the EoT scan is acquired.
Gradient Boosting Regressor (GBR) approach: Gradient Boosting Regressor was applied to standardized baseline PET-CT radiomic features from the 31 paired cases. Features were standardized using StandardScaler, and highly correlated features were removed before hyperparameter tuning. GridSearchCV optimized a grid of hyperparameters including number of estimators, learning rate, maximum depth, minimum samples split, and minimum samples leaf.
Prediction performance: The tuned GBR achieved an R-squared (R2) value of 0.86 plus/minus 0.09, indicating the model explained 86% of the variance in post-treatment TMTV. Mean Absolute Error (MAE) was 0.29 plus/minus 0.12, and Mean Absolute Percentage Error (MAPE) was 0.39 plus/minus 0.15, meaning predictions were on average about 39% off in relative terms. While the R2 value is notably high for a small retrospective dataset, the MAPE of 39% indicates that absolute accuracy at the individual patient level remains limited.
The authors emphasize that this experiment is a feasibility demonstration rather than a validated clinical tool. The relatively small training set of 31 cases means the GBR model is prone to overfitting, and the strong R2 may partly reflect overfit performance in cross-validation. The authors explicitly state that further investigation with larger training datasets is required before any clinical interpretation of this regression task is appropriate.
The most predictive delta radiomics features identified by the analysis span both PET and CT modalities and capture texture complexity rather than simple intensity or volume metrics. On the PET side, the top features included GLSZM Normalized Grey Level Non-uniformity, GLCM features (Difference Average, Correlation, Normalized Inverse Difference Moment), GLRLM Grey Level Non-uniformity, Intensity Histogram Root Mean Square and Uniformity, and local intensity-based features (Coefficient of Variation, Intensity Peak Discretized Volume Sought). On the CT side, leading features included GLSZM Zone Size Entropy, GLCM Sum Entropy and Zone Size Entropy, and local intensity-based features.
Spatial findings from RIM analysis: Radial Intensity Mean (RIM) analysis revealed that the innermost sub-regions of tumors (nearest the tumor core) were the most predictive in PET images. This finding aligns with known tumor biology: the core of aggressive lymphoma lesions often has the most heterogeneous metabolic activity, reflecting areas of necrosis, hypoxia, or proliferating clones. In CT images, a combination of sub-regions near the core and boundaries contributed, though individual boundary features were less predictive than their PET counterparts.
ML versus ICARE: Across both progression prediction and TTP analysis, conventional machine learning techniques (particularly RF and LDA) consistently outperformed ICARE. This contrasts with the original design intent of ICARE, which was developed specifically to provide robust predictions in small-sample clinical settings. One interpretation is that SFS feature selection combined with RF and LDA provides sufficient regularization for datasets of this size, reducing the specific advantage of ICARE's ensemble bootstrapping approach.
Generalizability to DLBCL: The authors tested whether their PMBCL-trained ICARE and ML models could predict progression in a separate cohort of 50 DLBCL cases based on EoT scans. Performance collapsed, with accuracy dropping to 0.34 and F1-score to 0.38 (near random for a 24% positive class). This failure of cross-disease generalization underscores that PMBCL and DLBCL have distinct imaging phenotypes despite sharing large B-cell histology. For multi-lesion DLBCL cases, the single-lesion PMBCL assumption (aggregating all lesions) is likely inappropriate; a weighted approach, or a focus on the largest or most metabolically active lesion, would be needed.
Small sample size and selection bias: The most significant limitation is the restricted cohort: only 31 patients had paired baseline and EoT scans suitable for delta radiomics analysis, and only 50 for single time-point EoT analysis. The paired cohort availability was determined by historical scanner resource constraints at BC Cancer (staging PET was not standard before 2011), introducing selection bias: patients from the earlier era were more likely to have EoT-only data, potentially representing a different treatment context and follow-up duration. The observed progression rates differ between cohorts (15.6% for paired versus 24.0% for EoT-only), which may partly reflect this era-driven selection rather than true biological difference.
Single-center retrospective design: All patients were treated at a single institution (BC Cancer), and imaging was acquired on two GE Healthcare scanners using a standardized protocol. Radiomic features are known to be sensitive to scanner hardware, acquisition parameters, and reconstruction algorithms. The generalizability of the radiomic signatures identified here to data from different scanners, institutions, or imaging protocols has not been tested. Multi-center prospective validation would be required before these findings could inform clinical practice.
Absence of interim PET analysis: The study examined only two time-points: pre-treatment baseline and end-of-treatment. Interim PET assessment (after 2 or 4 cycles of chemotherapy) is increasingly used for adaptive treatment strategies in DLBCL and has shown strong prognostic value. Delta radiomics computed between baseline and interim scans may provide earlier prognostic information and enable treatment modifications before completing the full R-CHOP course. This was explicitly cited as a limitation for future investigation.
Clinical variable integration: The prediction models relied exclusively on imaging features. Incorporation of clinical variables such as International Prognostic Index (IPI) components, LDH levels, circulating tumor DNA (ctDNA), and molecular/genetic characteristics could improve predictive performance. The authors specifically mention ctDNA as a promising complementary biomarker.
Path forward: Future validation should focus on larger multi-center PMBCL cohorts with prospective paired baseline and EoT scan collection, interim PET integration, and combination with clinical and molecular biomarkers. The PMBCL-specific nature of the radiomic signatures also suggests that separate model development and validation will be needed for each lymphoma subtype, given the failed generalization to DLBCL demonstrated in this study.