18F-FDG PET/CT Radiomics for Predicting Therapy Response in Primary Mediastinal B-Cell Lymphoma: A Bi-Centric Pilot Study

Abstract

Purpose: This bi-centric pilot study investigates the predictive value of pre-treatment [¹⁸F]FDG PET/CT radiomics for assessing therapy response in primary mediastinal B-cell lymphoma (PMBCL).

Methods: All PMBCL patients underwent PET/CT with [¹⁸F]FDG between January 2011 and January 2022 at Policlinico Tor Vergata University Hospital of Rome (70% training and 30% internal validation cohort) and Sant'Anna University Hospital of Ferrara (external validation cohort). The Deauville score (DS) was used as a predictor of therapy response (DS1-DS3 vs. DS4/DS5). A total of 121 quantitative radiomics features (RFts) were extracted from manually segmented volumes of interest (VOIs) in PET and CT images, according to IBSI. ComBat harmonization was applied to correct the center variability of features, followed by class balancing with SMOTE. Two machine learning (ML) prediction models, the PET model and the CT model, were independently developed using robust RFts. For each ML model, two different algorithms were trained (i.e., Random Forest, RF, and Support Vector Machine, SVM) using 10-fold cross validation, tested on the internal/external validation set. Receiver operating characteristic (ROC) curves, area under the curve (AUC), classification accuracy (CA), precision (Prec), sensitivity (Sen), specificity (Spec), true positive (TP) scores, and true negative (TN) scores were computed.

Results: The entire dataset was composed of 29 samples for the Rome cohort (23 from D1-D3 and 6 from D4/D5) and 9 samples for the Ferrara cohort (4 from D1-D3 and 5 from D4/D5). A total of 27 RFts were identified as robust for each imaging modality. Both the CT and PET models effectively predicted the Deauville score. The performance metrics of the best classifier (SVM) for the CT and PET models in external validation were AUC = 0.75/0.80, CA = 0.85/0.77, Prec = 0.97/0.67, Sen = 0.60/0.80, Spec = 0.98/0.75, TP = 75.0%/66.7%, and TN = 77.8%/85.7%, respectively.

Conclusions: ML models trained on [¹⁸F]FDG PET/CT radiomic features in PMBLC patients could predict the Deauville score.

Keywords: 18F-FDG; PET/CT; artificial intelligence; machine learning; primitive mediastinal B-cell lymphoma; radiomics.

Figure 1

Flow-chart of patient selection. A total of 38 PMBCL patients undergoing PET/CT at baseline and after therapy were enrolled from two centers: 29 in the training cohort and 9 in the validation cohort.

Figure 2

A case of PMBCL in a young woman (28 years old) from the University Hospital of Ferrara. Maximum intensity projection (A)—MIP—of baseline [¹⁸F]FDG PET/CT shows a large area of intense [¹⁸F]FDG uptake corresponding, in trans-axial fused PET/CT images (B), to a bulky mass of the anterior mediastinum. The image shows peripherical uptake with a “cold” core, corresponding to central necrosis. End-of-treatment [¹⁸F]FDG PET/CT shows a complete metabolic response (DS = 1) both in trans-axial fused PET/CT images (C) and at MIP (D). The patient did not experience disease relapse at follow-up (64 months).

Figure 3

ROC curves of the CT (a) and PET (b) models with both learners (SVM and RF) in internal validation.

Figure 4

ROC curves of both internal and external validation of CT (a) and PET (b) models.

Figure 5

ROC curve comparison (left) and radial plot of the performance scores (right) of SVM learners for CT and PET models.