Next Article in Journal
Phase Ia/Ib Study of Afatinib with Capecitabine in Patients with Refractory Solid Tumors and Pancreaticobiliary Cancers
Previous Article in Journal
Comparative Analysis of Cold Versus Thermal Dissection in Nerve-Sparing Robot-Assisted Radical Prostatectomy
Previous Article in Special Issue
Applicability of Radiomics for Differentiation of Pancreatic Adenocarcinoma from Healthy Tissue of Pancreas by Using Magnetic Resonance Imaging and Machine Learning
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

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

1
Hematology, Department of Biomedicine and Prevention, University of Rome “Tor Vergata”, 00133 Rome, Italy
2
Medical Physics Unit, University Hospital of Ferrara, Via Aldo Moro 8, 44124 Ferrara, Italy
3
Department of Translational Medicine, University of Ferrara, 44124 Ferrara, Italy
4
Nuclear Medicine Unit, Department of Onco-Hematology, Fondazione PTV Policlinico Tor Vergata University Hospital, 00133 Rome, Italy
5
Hematology Unit, University of Ferrara, 44121 Ferrara, Italy
6
Department of Biomedicine and Prevention, University of Rome “Tor Vergata”, 00133 Rome, Italy
7
Nuclear Medicine Unit, Onco-Hematology Department, University Hospital of Ferrara, 44124 Ferrara, Italy
*
Author to whom correspondence should be addressed.
These authors contributed equally to this work.
Cancers 2025, 17(11), 1827; https://doi.org/10.3390/cancers17111827
Submission received: 23 April 2025 / Revised: 26 May 2025 / Accepted: 28 May 2025 / Published: 30 May 2025
(This article belongs to the Special Issue Radiomics in Cancer Imaging: Theory and Applications in Solid Tumours)

Simple Summary

This bi-centric pilot study investigates the predictive value of pre-treatment [18F]FDG PET/CT radiomics for therapy response in primary mediastinal B-cell lymphoma (PMBCL). Patients from two Italian centers underwent baseline PET/CT, and the Deauville score (DS) was used to define response (DS1–3 vs. DS4–5). Radiomic features (RFts) were extracted from manually segmented PET and CT images and harmonized across centers. Two machine learning models (PET and CT) were trained with Random Forest and Support Vector Machine algorithms. In external validation, the best-performing SVM classifier showed AUCs of 0.80 (PET) and 0.75 (CT), with accuracies of 77% and 85%, respectively. Both models demonstrated strong specificity and precision. These results suggest that PET/CT radiomics-based ML models may assist in predicting treatment response in PMBCL patients.

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.

1. Introduction

Primary mediastinal B-cell lymphoma (PMBCL) accounts for approximately 2–3% of all non-Hodgkin lymphomas (NHLs) and mostly occurs in young individuals (median age of 35) [1]. It is a distinct entity in the current WHO classification and is characterized by specific clinicopathological and molecular features, with a gene expression profile partially overlapping with nodular sclerosis classical Hodgkin lymphoma [2,3,4]. Most patients present with a bulky anterior–superior mediastinal mass, exhibiting compressive symptoms like dry cough, dyspnea, or chest tightness. Mediastinal syndrome with jugular vein distension, thoracic wall venous reticulation, and “cape” edema, associated with superior vena cava thrombosis/obstruction, may occur. The disease typically presents as localized (Ann Arbor stages I–II) [5,6].
Histologically, it presents as a diffuse infiltrate of medium-to-large cells with clear cytoplasm, prominent nucleoli, and marked fibrosis. Tumor cells express B-cell antigens (CD19, CD20, CD22, and CD79A), with 80% showing weak to moderate CD30 positivity [7,8]. First-line therapeutic approaches may include various anthracycline-based regimens, such as dose-adjusted EPOCH-R (etoposide, prednisone, vincristine, cyclophosphamide, doxorubicin, and rituximab), R-CHOP14 (rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone every 14 days), and R-MACOPB (rituximab, methotrexate, doxorubicin, cyclophoshamide, vincristine prednisone, and bleomycin), all of which have demonstrated efficacy with the exception of R-CHOP21 (rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone every 21 days), which appears less effective in terms of response and exhibits favorable response rates compared to other types of NHL [9,10,11,12,13,14]. Second-line chemotherapy yields poor outcomes in relapsed cases [8,15]. Consolidation radiotherapy following first-line treatment was previously considered the standard of care [3]. While this approach showed promising results, it was not without medium- and long-term side effects, including pulmonary and cardiac complications, as well as secondary malignancies [16,17,18]. Currently, the radiotherapy approach is evolving, with the recent literature showing favorable outcomes in a radiation-free strategy for patients who achieve complete remission after treatment [19].
[18F]Fluorodeoxyglucose ([18F]FDG) PET/CT, as in all aggressive lymphomas, is the gold standard for staging, interim evaluation, and assessment at the end-of-treatment (EoT). For treatment response, the guidelines recommend using the five-point Deauville score (DS) [20,21]. The IELSG-26 study reported a high negative predictive value of PET at the end of treatment but a low positive predictive value. Likewise, an initial study by Melani et al. suggested that a negative EoT [18F]FDG PET scan is a strong predictor of cure, whereas a single positive EoT FDG-PET scan does not reliably indicate treatment failure [22]. Furthermore, the same study [22] reported that serial follow-up [18F]FDG-PET imaging can effectively differentiate between residual disease and post-treatment inflammatory changes, aiding in the identification of patients who may benefit from additional radiotherapy. Nevertheless, the interpretation of EoT PET/CT can present diagnostic challenges, as the heightened metabolic activity observed in residual mediastinal masses may yield spurious positive results, thereby complicating the formulation of optimal clinical management strategies. The interpretation of a DS of 4 remains frequently challenging, often requiring a new biopsy of the residual mediastinal mass to differentiate between inflammatory processes and treatment non-response [23]. Nonetheless, a negative [18F]FDG PET/CT at the end of treatment remains a well-established prognostic marker of a favorable outcome in clinical practice [22].
Radiomics, an emerging field in medical imaging, offers a promising solution to this challenge by extracting quantitative features from PET/CT scans that reflect tumor heterogeneity, metabolic activity, and the spatial distribution of [18F]FDG uptake [24]. Unlike conventional imaging interpretation, which relies on visual assessment, radiomics enables a more detailed and objective analysis of tumor characteristics, potentially improving response assessment accuracy and reducing unnecessary interventions [25,26]. Given the relatively indolent nature of residual PMBCL masses and the risk of overtreatment, integrating radiomic biomarkers into clinical practice could help refine post-treatment strategies and minimize unnecessary radiotherapy in patients who have already achieved metabolic remission.
Despite its potential, research on the role of [18F]FDG PET-based radiomics in PMBCL remains limited, partly due to the relatively low incidence of this oncological condition [27,28]. The primary objective of this pilot study is to evaluate the predictive value of pre-treatment [18F]FDG PET/CT radiomics in assessing treatment response in PMBCL. Specifically, this study aims to determine whether radiomic analysis can differentiate between patients who achieve a complete metabolic response (DS 1–3) and those with persistent or progressive disease (DS 4/5) at the end of therapy.

2. Materials and Methods

2.1. Study Design

This retrospective, Italian, bi-centric study included newly diagnosed patients with PMBCL who were treated with rituximab (R) and anthracycline-based therapeutic regimens and underwent [18F]FDG PET/CT between January 2011 and January 2022 at Policlinico Tor Vergata University Hospital of Rome (training and internal validation cohort) and Sant’Anna University Hospital of Ferrara (external validation cohort).
The inclusion criteria were as follows: (a) age ≥ 18 years; (b) histologically confirmed diagnosis of PMBCL; (c) baseline [18F]FDG PET/CT performed prior to the initiation of systemic therapies; (d) availability of EoT PET/CT; (e) a follow-up period of at least 24 months; and (f) a complete medical history. Conversely, exclusion criteria included (a) concomitant diagnosis of other oncological conditions; (b) unavailable EoT PET/CT; and (c) incomplete medical history or a follow-up period of less than 24 months.
For each patient enrolled, the following clinical data were collected: gender, age at diagnosis, disease stage, presence of B symptoms, Eastern Cooperative Oncology Group (ECOG) performance status, lactate dehydrogenase (LDH) levels, International Prognostic Index (IPI) categorization, presence of extranodal involvement, type of R-based treatment, whether consolidative radiotherapy was administered, whether disease progression occurred, and patient status at the last follow-up (progressive, alive with no evidence of disease).
The endpoint of this study was to assess whether [18F]FDG PET/CT-based radiomics, evaluated on pre-treatment scans, could predict EoT PET/CT results. For this purpose, at the EoT, patients with Deauville scores of DS1–3 were classified as having achieved a complete metabolic response, while those with DS4/DS5 were categorized as having a partial metabolic response or progressive disease (D5—new lesions or progression of existing ones). Notably, in this study, the determination of relapse or progression was based on information obtained from follow-up data.

2.2. [18F]FDG PET/CT Imaging

All patients underwent PET/CT with [18F]FDG, which was carried out according to the current imaging guidelines [29]. Two distinct scanners were used: each participating center performed PET/CT acquisition and reconstruction according to a standardized protocol detailed in previously published papers [30,31] and reported in Supplementary Materials.
In each center, PET/CT image analysis was performed by two experienced nuclear medicine physicians using Advantage 4.7 software (GE HealthCare, Wisconsin, MI, USA) at Tor Vergata University Hospital and the Syngo.via workstation (Siemens Healthineers, Erlangen, Germany) at Sant’Anna University Hospital of Ferrara. Any lesion exhibiting tracer uptake higher than the background and not categorized as physiological was considered potentially pathological. For each patient, the sites and number of lesions with pathological uptake were documented. The primary mediastinal mass and, when applicable, pathological lymph nodes were identified on PET images. Circular regions of interest (ROIs) were manually drawn and automatically converted into 3D volumes of interest (VOIs) by the software, as shown in Figure S1. The most common semiquantitative parameters were collected, including maximum and mean Standardized Uptake Values (SUVmax and SUVmean, respectively), metabolic tumor volume (MTV), and total lesion glycolysis (TLG). In each center, VOI segmentation of the mediastinal mass and the most significant pathological lymph nodes was manually performed on [18F]FDG PET/CT images by two expert nuclear medicine physicians using a 40% SUVmax threshold, employing the aforementioned software.

2.3. Feature Extraction

Quantitative radiomics features (RFts) were extracted from each VOI in PET and CT images. The extraction process was performed separately using the Radiomics package and the 3D Slicer image computing platform, in accordance with IBSI standardization [32].
A total of 121 RFts were obtained from each segmented VOI in both imaging modalities. These features were classified into different categories: 14 RFts corresponded to the original image and mask, 14 to the Shape (3D) class, 18 to first-order intensity statistics, 24 to the Gray Level Co-Occurrence Matrix (GLCM), 16 to the Gray Level Run Length Matrix (GLRLM), 16 to the Gray Level Size Zone Matrix (GLSZM), 14 to the Gray Level Dependence Matrix (GLDM), and 5 to the Neighboring Gray Tone Difference Matrix (NGTDM).

2.4. Data Preprocessing and Balancing Strategy

In order to correct the center variability of features, which can result from differences in imaging protocols or equipment across different centers, ComBat harmonization was applied [33]. Subsequently, we employed the Synthetic Minority Oversampling Technique (SMOTE) to generate synthetic samples, aiming to achieve a more balanced class distribution [34].
The cohort from Policlinico Tor Vergata University Hospital was then split into 70% for model training and 30% for internal validation. External validation was performed using a dataset from Sant’Anna University Hospital. No additional normalization or transformations were applied to the data.

2.5. Statistical Analysis and Model Construction

A filter-based feature selection algorithm, implemented using a custom Python (Version 3.8) script, was used to identify robust and non-redundant RFts. To determine the most reproducible RFts, a significance threshold of p < 0.05 in the Wilcoxon–Mann–Whitney U test was applied.
Two machine learning (ML) prediction models, the PET model and the CT model, were independently developed. For each ML model, two different algorithms were trained: Random Forest (RF) and Support Vector Machine (SVM). The Orange data-mining platform was used for training, as well as for internal and external validation of both ML models based on PET and CT robust RFts.
The RF and SVM models were trained using 10-fold cross-validation (10FoldCV), tested on the internal validation set, and externally validated on the respective dataset.
Receiver operating characteristic (ROC) curves were plotted, and the area under the ROC curve (AUC), classification accuracy (CA), precision, sensitivity, specificity, true positive (TP) scores, and true negative (TN) scores were computed. These performance metrics were used to determine the best-performing algorithm for the PET and CT models.

3. Results

Initially, 50 potentially eligible patients were identified in the training cohort. However, 21 patients were excluded due to incomplete documentation (n = 3) or non-recoverable/damaged diagnostic exams (n = 18), leaving 29 patients in the training cohort. In the validation cohort, nine patients were identified, all of whom met the inclusion criteria (Figure 1). Ultimately, 38 patients were enrolled in the study, with 29 in the training cohort and 9 in the validation cohort. The demographic and clinicopathological characteristics of the selected patients are summarized in Table 1.

3.1. Cohorts’ Composition, Characteristics, and Outcome

The median age was slightly higher in the validation cohort (39.5 years, range 25–58) than in the training cohort (35 years, range 18–58). The distribution of the IPI scores varied, with a greater proportion of patients in the training cohort presenting with lower IPI scores. Notably, B symptoms were absent in the validation cohort but present in nearly one-third of the training cohort. Most patients in both groups had a good performance status (ECOG 0–1). Bulky disease was frequent in both cohorts, while extranodal involvement was observed only in the training cohort. Regarding treatment, the training cohort received a variety of anthracycline-based regimens, predominantly R-MACOP-B, while all patients in the validation cohort were treated with RCHOP.
At the end of treatment, the majority of patients in the training cohort (23 out of 29) achieved a DS of 1–3, whereas the validation cohort had a more evenly distributed response, with four patients classified as responders and five as non-responders (DS 4–5). A representative case of a patient with a large, hypermetabolic mediastinal mass at pre-treatment PET/CT, showing a complete metabolic response (DS1) after therapy, is illustrated in Figure 2. Consolidative radiotherapy was administered more frequently in the training cohort, and relapse occurred in only two patients. At the last follow-up, all patients in both cohorts were alive, with no evidence of disease, and the median follow-up time was 4.66 years in the training cohort and 3.64 years in the validation cohort.

3.2. Radiomic Analysis and Model Building

After applying the class balancing process using the SMOTE methodology, the entire dataset was composed of 47 samples for the training cohort (25 from DS1–DS3 and 22 from DS4/DS5) and 13 samples for the validation cohort (5 from DS1–DS3 and 8 from DS4/DS5). Among the 122 RFts extracted from CT and PET, 27 were identified as robust for each imaging modality. The full list of selected RFts is reported in Supplementary Material (Tables S1 and S2).
The performance of 10FoldCV across both learners demonstrated an AUC > 0.90 and a CA > 0.80 for the CT and PET models, respectively. During the training phase, the CT model reached the best performance, with the SVM classifier achieving an AUC > 0.97 and a CA > 0.88. In the PET model, the SVM classifier outperformed the RF classifier, reaching an AUC of 0.86 and a CA of 0.86. The performance of the two models, with both learners, during the training phase is shown in Table 2.
The internal validation, performed on the remaining 30% of the training dataset, yielded performance results consistent with those observed during training. In the CT model, the SVM classifier achieved an AUC of 0.90, with a TP of 80% and a TN of 75%.
The external validation, carried out on the validation cohort, maintained good classification capabilities for the models built. The set of performance scores for the validation processes are presented in Table 3.
The ROC curves for the CT and PET models with both learners (SVM and RF) are shown in Figure 3. For the best-performing learner, identified as the SVM, the ROC curves for both internal and external validation are plotted (Figure 4).
Additionally, the ROC curves obtained during the training phase are provided in the Supplementary Materials (Figure S2).
Finally, a comparison of the results was made, and a radial plot was created to visually compare the performance scores of SVM learners for each of the two ML models trained. Figure 5 presents a graphical comparison of the performance of the SVM, which emerged as the best learner in terms of performance across both models.

4. Discussion

[18F]FDG PET plays a crucial role in the management of PMBCL, offering metabolic insights that complement anatomical imaging. [18F]FDG PET/CT is not only essential for initial staging but also serves as a key tool in response evaluation, guiding subsequent therapeutic strategies. A significant challenge, especially after treatment, is the prevalence of false-positive results. Patients may present with a reduction in the size of the mediastinal mass, with residual focal findings evidenced at end of treatment [18F]FDG PET/CT, corresponding to a high DS (i.e., DS 4–5). However, in many cases, this does not reflect active disease, but rather necrotic inflammatory tissue due to chemotherapy, as confirmed by re-biopsy of the residual mediastinal mass. To date, there are no effective prognostic scores for PMBCL to understand the outcome of the disease at onset or drive therapeutic choices. The predictive value of [18F]FDG PET has been demonstrated in different studies; in particular, a negative PET scan at the end of chemoimmunotherapy is associated with prolonged progression-free survival (PFS) [35].
First-line treatment of PMBCL yields a high probability of cure. Regimens typically involve the use of rituximab in combination with intensive anthracycline-based chemotherapy. However, the prognosis is significantly impaired in patients who do not achieve a complete response or relapse after initial therapy [8,15]. Currently, also due to the rarity of the incidence of PMBCL, there is no standardized first-line treatment regimen, and there are no effective prospective randomized trials able to define the best chemoimmunotherapy treatment in this disease [9,36,37].
Initially, the standard treatment strategy included anthracycline-based chemotherapy with rituximab, as a CHOP-like regimen, followed by consolidation radiotherapy [38]. Dunleavy et al., in a prospective phase II study, demonstrated the efficacy of chemo-immunotherapy according to the DA-EPOCH-R regimen with omission of radiotherapy in patients with PMBCL [9]. Limited and conflicting data exist regarding the comparative effectiveness of different chemotherapy regimens. While some studies suggest DA-EPOCH-R superiority over R-CHOP, others report no significant outcome difference [37,39,40]. The DA-EPOCH-R regimen, while effective and potentially sparing patients from radiotherapy, presents a higher toxicity profile. This includes increased hospitalization, the need for central venous access (often difficult to find due to the presence of mediastinal bulky masses present at onset that expose patients to increased risk of venous thrombosis), and the increased use of granulocyte growth factors (G-CSFs) for more frequent and longer-lasting hematologic toxicities than the other regimens used in frontline PMBCL [36].
Recently, the IELSG37 study has revolutionized the therapeutic approach in patients, with PMBCL leading to the omission of radiotherapy in patients with complete metabolic response (DS1–3) [19,41]. A sub-analysis of this study compared the efficacy of different first-line chemo-immunotherapy regimens (R-CHOP 14, R-CHOP 21, R V/MACOP-B, DA-EPOCH-R), and, although DA-EPOCH-R has not demonstrated a statistically significant advantage, there is a trend towards efficacy in its favor compared to other therapeutic regimens [14]. It has to be emphasized that the choice of treatment often depends on the physician’s decision and does not reflect standard parameters. In order to identify high-risk patients, there are no prognostic scores to guide clinical practice in the treatment decision, and the IPI, used mainly in DLBCL, does not guide the choice of treatment and is not specific to PMBCL. In this context, it would be important to more effectively identify high-risk patients who might benefit from more intensive treatment, even facing an increased toxicity [38].
Given the role of [18F]FDG PET/CT in patients with PMBCL, recent studies have focused on the incorporation of quantitative functional parameters, such as MTV and TLG, to identify patients at high risk of treatment non-response who might benefit from more intensive treatments such as DA-EPOCH-R. A correlation was demonstrated between [18F]FDG PET-derived parameters, such as MTV and TLG and patient outcome in terms of OS and PFS [28,35,42]. One of the emerging findings in PMBCL is the role of baseline metabolic heterogeneity (MH) in predicting treatment outcomes. The presence of intratumoral heterogeneity, reflected in parameters such as textural features on PET scans, has been associated with an increased risk of treatment resistance and disease recurrence [42]. Tumors with high metabolic heterogeneity may harbor more aggressive clones, necessitating closer monitoring and potentially more intensive therapeutic approaches. However, the calculation of MH does not follow a standardized approach, which also leads to a certain variability in its clinical implication [43,44,45]. Advanced PET metrics may help mitigate these issues by offering more standardized cutoffs for residual disease assessment [46]. In this perspective, AI-driven PET analysis holds the potential to improve diagnostic accuracy by reducing interobserver variability and enabling automated risk stratification [47,48]. The integration of radiomic analysis into PET interpretation could refine prognostic assessments, enabling personalized treatment adjustments. Preliminary studies suggest that radiomic analysis can capture subtle variations in metabolic activity across voxels, reflecting fluctuations in tumor composition that are not immediately interpretable from a clinical standpoint [49]. However, these radiomic features have been found to strongly correlate with key biological tumor characteristics and have demonstrated significant prognostic value, being closely linked to treatment response and survival outcomes [48,50,51].
The most commonly used model building strategy involves splitting the training sample into two parts for training and testing, referred to as internal validation. The trained learners are the Random Forest (RF) and Support Vector Machine (SVM) algorithms. The performance of the two learners has no significant differences; both perform well, but the SVM demonstrates better adaptation to the data. This results in superior performance during 10FoldCV (AUC > 0.85, CA > 0.86, PRE > 0.85, SEN > 0.84, SPE > 0.85). This performance is replicated in both internal validation (AUC > 0.84, CA > 0.85, PRE > 0.85, SEN > 0.84, SPE > 0.86) and external validation (AUC > 0.75 CA > 0.77, PRE > 0.67, SEN > 0.60, SPE > 0.75). The evaluations of true positive (TP) and true negative (TN) scores are very interesting and significant, showing that there is a strong predictive power for the DS when analyzing the RFts of [18F]FDG PET/CT. In particular, the best scores are 77.80% for the TN with the CT model and 85.70% with the PET model. The statistically robust process of feature selection was confirmed by the stability of the two models. In fact, both the CT and PET models showed potential in predicting DS in PMBCL patients.
A common limitation often highlighted in radiomics studies is the lack of an external database, which prevents proper data validation. However, in our study, access to a database from another center allowed us to validate our models. Future developments of the project include validation on a larger cohort and prospective studies. In spite of its limited cohort of patients, our study offers elements that add value to the field, including a rigorous feature selection process, which effectively identifies the most representative radiomic features set for the binary problem under investigation. The use of harmonization techniques helps to mitigate the impact of different scanners, directly improving model performance in both the training and validation phases. As discussed by Orlhac and colleagues [52], harmonization in medical imaging can also be considered as a form of domain adaptation, where the goal is to produce images belonging to a single domain. It is important to highlight that the validation dataset, even if it is smaller than the training dataset, serves as a direct validation of the effective training of the proposed machine learning models. The use of an external unknown validation dataset in predictive ML models is essential to ensure reliability and to move closer to standardization in clinical practice. Finally, when comparing internal and external validation cohorts, serum LDH > 225 U/L was observed in all patients from the external validation cohort and in two-thirds of patients from the training cohort. Since increased LDH reflects greater and more aggressive neoplastic mass, we cannot exclude the possibility that the observed imbalance in baseline LDH—reflecting differences in tumor burden and aggressiveness between the training and external validation cohorts—contributed to the slightly lower true positive rates in the external set; however, our use of SMOTE to rebalance responders and non-responders likely mitigated this artifact. In this respect, the increasing sophistication of imaging techniques and the growing use of artificial intelligence and predictive models further highlight the need for a new standardization framework in multi- or bi-centric studies on PMBCL. This pilot study serves as a gateway to advanced clinical research in PET/CT and radiomics.
Radiomics, as observed in DLBCL and HL, may aid PMBCL risk stratification and future treatment decisions. Integrating biological, clinical, histological, and radiomic parameters could produce effective prognostic models for clinical use [40,53,54,55].

5. Conclusions

In conclusion, our pilot study suggests that radiomic features extracted from pre-treatment [18F]FDG PET/CT can effectively predict therapy response in PMBCL, as reflected by the DS. The application of machine learning models, particularly the SVM classifier, showed promising accuracy and robustness across both internal and external validations. Despite the limited sample size, these findings underscore the potential of integrating radiomics into clinical decision-making to guide personalized treatment strategies. Further large-scale, prospective studies are warranted to validate these preliminary results and refine prognostic models.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/cancers17111827/s1, Figure S1: 25-year-old man with newly diagnosed primary mediastinal B-cell lymphoma. (A) Whole-body PET shows an extensive area of increased tracer uptake in the mediastinum. Axial (B) and coronal (C) fused PET/CT images demonstrate lesion segmentation using a 40% SUVmax threshold (yellow contour); Figure S2: ROC curves obtained during the training phase for the PET and CT-based models; Table S1: CT model selected robust radiomic features; Table S2: PET model selected robust radiomic features.

Author Contributions

Conceptualization, L.F., F.E., L.M. and L.U.; methodology, R.G., L.M. and G.S.; software, L.M. and G.S.; validation, L.M. and G.S.; data curation F.E., L.F., L.U. and S.A.; writing—original draft preparation, L.F., F.E., L.M. and L.U.; writing—review and editing, L.D.M., A.V., M.P., N.U., A.C. (Agostino Chiaravalloti), A.C. (Antonio Cuneo) and M.B. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

The study was conducted according to the Declaration of Helsinki, Good Clinical Practice, and local ethical regulations. The local ethical committee of each participating center approved the study (protocol n. 80.25 of CET Lazio Area 2 for Tor Vergata University Hospital and protocol CET-AVEC 520-2024-Oss-AOUFe for Ferrara University Hospital).

Informed Consent Statement

Informed consent was obtained from all subjects involved in the study.

Data Availability Statement

Datasets generated and analyzed during the study are available upon reasonable request to the corresponding author.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Yu, Y.; Dong, X.; Tu, M.; Wang, H. Primary Mediastinal Large B Cell Lymphoma. Thorac. Cancer 2021, 12, 2831–2837. [Google Scholar] [CrossRef] [PubMed]
  2. Rosenwald, A.; Wright, G.; Leroy, K.; Yu, X.; Gaulard, P.; Gascoyne, R.D.; Chan, W.C.; Zhao, T.; Haioun, C.; Greiner, T.C.; et al. Molecular Diagnosis of Primary Mediastinal B Cell Lymphoma Identifies a Clinically Favorable Subgroup of Diffuse Large B Cell Lymphoma Related to Hodgkin Lymphoma. J. Exp. Med. 2003, 198, 851–862. [Google Scholar] [CrossRef] [PubMed]
  3. Martelli, M.; Ferreri, A.; Di Rocco, A.; Ansuinelli, M.; Johnson, P.W.M. Primary Mediastinal Large B-Cell Lymphoma. Crit. Rev. Oncol. Hematol. 2017, 113, 318–327. [Google Scholar] [CrossRef]
  4. Alaggio, R.; Amador, C.; Anagnostopoulos, I.; Attygalle, A.D.; Araujo, I.B.D.O.; Berti, E.; Bhagat, G.; Borges, A.M.; Boyer, D.; Calaminici, M.; et al. The 5th Edition of the World Health Organization Classification of Haematolymphoid Tumours: Lymphoid Neoplasms. Leukemia 2022, 36, 1720–1748. [Google Scholar] [CrossRef]
  5. Johnson, P.W.M.; Davies, A.J. Primary Mediastinal B-Cell Lymphoma. Hematology 2008, 2008, 349–358. [Google Scholar] [CrossRef]
  6. Zinzani, P.L.; Martelli, M.; Poletti, V.; Vitolo, U.; Gobbi, P.G.; Chisesi, T.; Barosi, G.; Ferreri, A.J.M.; Marchetti, M.; Pimpinelli, N.; et al. Practice Guidelines for the Management of Extranodal Non-Hodgkin’s Lymphomas of Adult Non-Immunodeficient Patients. Part I: Primary Lung and Mediastinal Lymphomas. A Project of the Italian Society of Hematology, the Italian Society of Experimental Hematology and the Italian Group for Bone Marrow Transplantation. Haematologica 2008, 93, 1364–1371. [Google Scholar] [CrossRef] [PubMed]
  7. Yamamoto, W.; Nakamura, N.; Tomita, N.; Ishii, Y.; Takasaki, H.; Hashimoto, C.; Motomura, S.; Yamazaki, E.; Ohshima, R.; Numata, A.; et al. Clinicopathological Analysis of Mediastinal Large B-Cell Lymphoma and Classical Hodgkin Lymphoma of the Mediastinum. Leuk. Lymphoma 2013, 54, 967–972. [Google Scholar] [CrossRef]
  8. Donzelli, L.; Rocco, A.D.; Petrucci, L.; Martelli, M. Primary Mediastinal Large B-Cell Lymphoma: Biological Features, Clinical Characteristics and Current Treatment Strategies. Cancer Treat. Rev. 2025, 134, 102898. [Google Scholar] [CrossRef] [PubMed]
  9. Dunleavy, K.; Pittaluga, S.; Maeda, L.S.; Advani, R.; Chen, C.C.; Hessler, J.; Steinberg, S.M.; Grant, C.; Wright, G.; Varma, G.; et al. Dose-Adjusted EPOCH-Rituximab Therapy in Primary Mediastinal B-Cell Lymphoma. N. Engl. J. Med. 2013, 368, 1408–1416. [Google Scholar] [CrossRef]
  10. Stepanishyna, Y.; Skrypets, T.; Novosad, O.; Pastushenko, I.; Kryachok, I. DA-EPOCH-R vs. R-CHOP in Patients with Primary Mediastinal Large B-Cell Lymphoma: Results of Prospective Randomized Ukrainian Multicenter Study. Blood 2020, 136, 9. [Google Scholar] [CrossRef]
  11. Gleeson, M.; Hawkes, E.A.; Cunningham, D.; Chadwick, N.; Counsell, N.; Lawrie, A.; Jack, A.; Smith, P.; Mouncey, P.; Pocock, C.; et al. Rituximab, Cyclophosphamide, Doxorubicin, Vincristine and Prednisolone (R- CHOP) in the Management of Primary Mediastinal B-cell Lymphoma: A Subgroup Analysis of the UK NCRI R- CHOP 14 versus 21 Trial. Br. J. Haematol. 2016, 175, 668–672. [Google Scholar] [CrossRef]
  12. Malenda, A.; Kołkowska-Leśniak, A.; Puła, B.; Długosz-Danecka, M.; Chełstowska, M.; Końska, A.; Giza, A.; Lech-Marańda, E.; Jurczak, W.; Warzocha, K. Outcomes of Treatment with Dose-adjusted EPOCH-R or R-CHOP in Primary Mediastinal Large B-cell Lymphoma. Eur. J. Haematol. 2020, 104, 59–66. [Google Scholar] [CrossRef] [PubMed]
  13. Cook, M.R.; Williams, L.S.; Dorris, C.S.; Luo, Y.; Makambi, K.; Dunleavy, K. Improved Survival for Dose-Intensive Chemotherapy in Primary Mediastinal B-Cell Lymphoma: A Systematic Review and Meta-Analysis of 4068 Patients. Haematologica 2023, 109, 846–856. [Google Scholar] [CrossRef]
  14. Martelli, M.; Zucca, E.; Botto, B.; Kryachok, I.; Ceriani, L.; Balzarotti, M.; Tucci, A.; Cabras, M.G.; Zilioli, V.R.; Rusconi, C.; et al. Impact of Different Induction Regimens on the Outcome of Primary Mediastinal B Cell Lymphoma in the Prospective Ielsg 37 Trial. Hematol. Oncol. 2021, 39, hon.49_2879. [Google Scholar] [CrossRef]
  15. Vardhana, S.; Hamlin, P.A.; Yang, J.; Zelenetz, A.; Sauter, C.S.; Matasar, M.J.; Ni, A.; Yahalom, J.; Moskowitz, C.H. Outcomes of Relapsed and Refractory Primary Mediastinal (Thymic) Large B Cell Lymphoma Treated with Second-Line Therapy and Intent to Transplant. Biol. Blood Marrow Transplant. 2018, 24, 2133–2138. [Google Scholar] [CrossRef] [PubMed]
  16. Mazzola, R.; Giaj Levra, N.; Alongi, F. Radiation Dose-Response Relationship for Risk of Coronary Heart Disease in Survivors of Hodgkin Lymphoma. J. Clin. Oncol. 2016, 34, 2940–2941. [Google Scholar] [CrossRef]
  17. Van Leeuwen, F.E.; Ng, A.K. Long-Term Risk of Second Malignancy and Cardiovascular Disease after Hodgkin Lymphoma Treatment. Hematol. Am. Soc. Hematol. Educ. Program. 2016, 2016, 323–330. [Google Scholar] [CrossRef]
  18. Nassi, L.; De Sanctis, V.; Loseto, G.; Gerardi, C.; Allocati, E.; Ciavarella, S.; Minoia, C.; Guarini, A.; Bari, A. Second Cancers in Classical Hodgkin Lymphoma and Diffuse Large B-Cell Lymphoma: A Systematic Review by the Fondazione Italiana Linfomi. Cancers 2022, 14, 519. [Google Scholar] [CrossRef]
  19. Martelli, M.; Ceriani, L.; Ciccone, G.; Ricardi, U.; Kriachok, I.; Botto, B.; Balzarotti, M.; Tucci, A.; Usai, S.V.; Zilioli, V.R.; et al. Omission of Radiotherapy in Primary Mediastinal B-Cell Lymphoma: IELSG37 Trial Results. J. Clin. Oncol. 2024, 42, 4071–4083. [Google Scholar] [CrossRef]
  20. Martelli, M.; Ceriani, L.; Zucca, E.; Zinzani, P.L.; Ferreri, A.J.M.; Vitolo, U.; Stelitano, C.; Brusamolino, E.; Cabras, M.G.; Rigacci, L.; et al. [18F]Fluorodeoxyglucose Positron Emission Tomography Predicts Survival After Chemoimmunotherapy for Primary Mediastinal Large B-Cell Lymphoma: Results of the International Extranodal Lymphoma Study Group IELSG-26 Study. J. Clin. Oncol. 2014, 32, 1769–1775. [Google Scholar] [CrossRef]
  21. Cheson, B.D.; Fisher, R.I.; Barrington, S.F.; Cavalli, F.; Schwartz, L.H.; Zucca, E.; Lister, T.A. Recommendations for Initial Evaluation, Staging, and Response Assessment of Hodgkin and Non-Hodgkin Lymphoma: The Lugano Classification. J. Clin. Oncol. 2014, 32, 3059–3067. [Google Scholar] [CrossRef] [PubMed]
  22. Melani, C.; Advani, R.; Roschewski, M.; Walters, K.M.; Chen, C.C.; Baratto, L.; Ahlman, M.A.; Miljkovic, M.D.; Steinberg, S.M.; Lam, J.; et al. End-of-Treatment and Serial PET Imaging in Primary Mediastinal B-Cell Lymphoma Following Dose-Adjusted EPOCH-R: A Paradigm Shift in Clinical Decision Making. Haematologica 2018, 103, 1337–1344. [Google Scholar] [CrossRef] [PubMed]
  23. Adams, H.J.A.; Kwee, T.C. Unproven Value of End-of-Treatment and Serial Follow-up FDG-PET in Primary Mediastinal B-Cell Lymphoma. Haematologica 2018, 103, e380–e381. [Google Scholar] [CrossRef]
  24. Gillies, R.J.; Kinahan, P.E.; Hricak, H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016, 278, 563–577. [Google Scholar] [CrossRef] [PubMed]
  25. Albano, D.; Ravanelli, M.; Durmo, R.; Versari, A.; Filice, A.; Rizzo, A.; Racca, M.; Pizzuto, D.A.; Bertagna, F.; Annunziata, S. Semiquantitative 2-[18F]FDG PET/CT-Based Parameters Role in Lymphoma. Front. Med. 2024, 11, 1515040. [Google Scholar] [CrossRef]
  26. Filippi, L.; Ferrari, C.; Nuvoli, S.; Bianconi, F.; Donner, D.; Marongiu, A.; Mammucci, P.; Vultaggio, V.; Chierichetti, F.; Rubini, G.; et al. Pet-Radiomics in Lymphoma and Multiple Myeloma: Update of Current Literature. Clin. Transl. Imaging 2023, 12, 119–135. [Google Scholar] [CrossRef]
  27. Abenavoli, E.M.; Barbetti, M.; Linguanti, F.; Mungai, F.; Nassi, L.; Puccini, B.; Romano, I.; Sordi, B.; Santi, R.; Passeri, A.; et al. Characterization of Mediastinal Bulky Lymphomas with FDG-PET-Based Radiomics and Machine Learning Techniques. Cancers 2023, 15, 1931. [Google Scholar] [CrossRef]
  28. Yousefirizi, F.; Gowdy, C.; Klyuzhin, I.S.; Sabouri, M.; Tonseth, P.; Hayden, A.R.; Wilson, D.; Sehn, L.H.; Scott, D.W.; Steidl, C.; et al. Evaluating Outcome Prediction via Baseline, End-of-Treatment, and Delta Radiomics on PET-CT Images of Primary Mediastinal Large B-Cell Lymphoma. Cancers 2024, 16, 1090. [Google Scholar] [CrossRef]
  29. Boellaard, R.; Delgado-Bolton, R.; Oyen, W.J.G.; Giammarile, F.; Tatsch, K.; Eschner, W.; Verzijlbergen, F.J.; Barrington, S.F.; Pike, L.C.; Weber, W.A.; et al. FDG PET/CT: EANM Procedure Guidelines for Tumour Imaging: Version 2.0. Eur. J. Nucl. Med. Mol. Imaging 2015, 42, 328–354. [Google Scholar] [CrossRef]
  30. Filippi, L.; Proietti, I.; Morea, S.; Potenza, C. Concomitant Thyroiditis and Orchitis Induced by Immune Checkpoint Inhibitors Detected on [18F]FDG PET/CT. Egypt. J. Radiol. Nucl. Med. 2024, 55, 49. [Google Scholar] [CrossRef]
  31. Urso, L.; Panareo, S.; Castello, A.; Ambrosio, M.R.; Zatelli, M.C.; Caracciolo, M.; Tonini, E.; Valpiani, G.; Boschi, A.; Uccelli, L.; et al. Glucose Metabolism Modification Induced by Radioligand Therapy with [177Lu]Lu/[90Y]Y-DOTATOC in Advanced Neuroendocrine Neoplasms: A Prospective Pilot Study within FENET-2016 Trial. Pharmaceutics 2022, 14, 2009. [Google Scholar] [CrossRef] [PubMed]
  32. Fedorov, A.; Beichel, R.; Kalpathy-Cramer, J.; Finet, J.; Fillion-Robin, J.-C.; Pujol, S.; Bauer, C.; Jennings, D.; Fennessy, F.; Sonka, M.; et al. 3D Slicer as an Image Computing Platform for the Quantitative Imaging Network. Magn. Reson. Imaging 2012, 30, 1323–1341. [Google Scholar] [CrossRef]
  33. Behdenna, A.; Colange, M.; Haziza, J.; Gema, A.; Appé, G.; Azencott, C.-A.; Nordor, A. pyComBat, a Python Tool for Batch Effects Correction in High-Throughput Molecular Data Using Empirical Bayes Methods. BMC Bioinform. 2023, 24, 459. [Google Scholar] [CrossRef]
  34. Chawla, N.V.; Bowyer, K.W.; Hall, L.O.; Kegelmeyer, W.P. SMOTE: Synthetic Minority Over-Sampling Technique. J. Artif. Intell. Res. 2002, 16, 321–357. [Google Scholar] [CrossRef]
  35. Ceriani, L.; Milan, L.; Johnson, P.W.M.; Martelli, M.; Presilla, S.; Giovanella, L.; Zucca, E. Baseline PET Features to Predict Prognosis in Primary Mediastinal B Cell Lymphoma: A Comparative Analysis of Different Methods for Measuring Baseline Metabolic Tumour Volume. Eur. J. Nucl. Med. Mol. Imaging 2019, 46, 1334–1344. [Google Scholar] [CrossRef] [PubMed]
  36. Melani, C.; Wilson, W.H.; Roschewski, M. What Is the Standard of Care for Primary Mediastinal B-cell Lymphoma; R- CHOP or DA-EPOCH-R? Br. J. Haematol. 2019, 184, 836–838. [Google Scholar] [CrossRef]
  37. Shah, N.N.; Szabo, A.; Huntington, S.F.; Epperla, N.; Reddy, N.; Ganguly, S.; Vose, J.; Obiozor, C.; Faruqi, F.; Kovach, A.E.; et al. R- CHOP versus Dose-adjusted R- EPOCH in Frontline Management of Primary Mediastinal B-cell Lymphoma: A Multi-centre Analysis. Br. J. Haematol. 2018, 180, 534–544. [Google Scholar] [CrossRef]
  38. Liu, Y.; Jiang, J.; Liu, L.; Wang, Z.; Yu, B.; Xia, Z.; Zhang, Q.; Ji, D.; Liu, X.; Lv, F.; et al. Prognostic Significance of Clinical Characteristics and18 Fluorodeoxyglucose-Positron Emission Tomography/Computed Tomography Quantitative Parameters in Patients with Primary Mediastinal B-Cell Lymphoma. J. Int. Med. Res. 2022, 50, 03000605211063027. [Google Scholar] [CrossRef] [PubMed]
  39. Chan, E.H.L.; Koh, L.P.; Lee, J.; De Mel, S.; Jeyasekharan, A.; Liu, X.; Tang, T.; Lim, S.T.; Tao, M.; Quek, R.; et al. Real World Experience of R-CHOP with or without Consolidative Radiotherapy vs. DA-EPOCH-R in the First-line Treatment of Primary Mediastinal B-cell Lymphoma. Cancer Med. 2019, 8, 4626–4632. [Google Scholar] [CrossRef]
  40. Bartlett, N.L.; Wilson, W.H.; Jung, S.-H.; Hsi, E.D.; Maurer, M.J.; Pederson, L.D.; Polley, M.-Y.C.; Pitcher, B.N.; Cheson, B.D.; Kahl, B.S.; et al. Dose-Adjusted EPOCH-R Compared With R-CHOP as Frontline Therapy for Diffuse Large B-Cell Lymphoma: Clinical Outcomes of the Phase III Intergroup Trial Alliance/CALGB 50303. J. Clin. Oncol. 2019, 37, 1790–1799. [Google Scholar] [CrossRef]
  41. Ceriani, L.; Milan, L.; Pirosa, M.C.; Martelli, M.; Ruberto, T.; Cascione, L.; Johnson, P.W.M.; Davies, A.J.; Ciccone, G.; Zucca, E. PET-Based Risk Stratification in Primary Mediastinal B-Cell Lymphoma: A Comparative Analysis of Different Segmentation Methods in the IELSG37 Trial Patient Cohort. J. Nucl. Med. 2025, 66, 209–214. [Google Scholar] [CrossRef] [PubMed]
  42. Ceriani, L.; Milan, L.; Martelli, M.; Ferreri, A.J.M.; Cascione, L.; Zinzani, P.L.; Di Rocco, A.; Conconi, A.; Stathis, A.; Cavalli, F.; et al. Metabolic Heterogeneity on Baseline 18FDG-PET/CT Scan Is a Predictor of Outcome in Primary Mediastinal B-Cell Lymphoma. Blood 2018, 132, 179–186. [Google Scholar] [CrossRef]
  43. Senjo, H.; Hirata, K.; Izumiyama, K.; Minauchi, K.; Tsukamoto, E.; Itoh, K.; Kanaya, M.; Mori, A.; Ota, S.; Hashimoto, D.; et al. High Metabolic Heterogeneity on Baseline 18FDG-PET/CT Scan as a Poor Prognostic Factor for Newly Diagnosed Diffuse Large B-Cell Lymphoma. Blood Adv. 2020, 4, 2286–2296. [Google Scholar] [CrossRef] [PubMed]
  44. Liu, J.; Ren, Q.; Xiao, H.; Li, S.; Zheng, L.; Yang, X.; Feng, L.; Zhou, Z.; Wang, H.; Yang, J.; et al. Whole-Tumoral Metabolic Heterogeneity in 18F-FDG PET/CT Is a Novel Prognostic Marker for Neuroblastoma. Cancer Imaging 2024, 24, 72. [Google Scholar] [CrossRef] [PubMed]
  45. Sanli, Y.; Leake, J.; Odu, A.; Xi, Y.; Subramaniam, R.M. Tumor Heterogeneity on FDG PET/CT and Immunotherapy: An Imaging Biomarker for Predicting Treatment Response in Patients With Metastatic Melanoma. AJR Am. J. Roentgenol. 2019, 212, 1318–1326. [Google Scholar] [CrossRef]
  46. Kostakoglu, L.; Chauvie, S. PET-Derived Quantitative Metrics for Response and Prognosis in Lymphoma. PET Clin. 2019, 14, 317–329. [Google Scholar] [CrossRef]
  47. Cairns, J.; Frood, R.; Patel, C.; Scarsbrook, A. The Role of AI in Lymphoma: An Update. Semin. Nucl. Med. 2025, 55, 377–386. [Google Scholar] [CrossRef]
  48. Evangelista, L.; Fiz, F.; Laudicella, R.; Bianconi, F.; Castello, A.; Guglielmo, P.; Liberini, V.; Manco, L.; Frantellizzi, V.; Giordano, A.; et al. PET Radiomics and Response to Immunotherapy in Lung Cancer: A Systematic Review of the Literature. Cancers 2023, 15, 3258. [Google Scholar] [CrossRef]
  49. Henry, T.; Sun, R.; Lerousseau, M.; Estienne, T.; Robert, C.; Besse, B.; Robert, C.; Paragios, N.; Deutsch, E. Investigation of Radiomics Based Intra-Patient Inter-Tumor Heterogeneity and the Impact of Tumor Subsampling Strategies. Sci. Rep. 2022, 12, 17244. [Google Scholar] [CrossRef]
  50. Veziroglu, E.M.; Farhadi, F.; Hasani, N.; Nikpanah, M.; Roschewski, M.; Summers, R.M.; Saboury, B. Role of Artificial Intelligence in PET/CT Imaging for Management of Lymphoma. Semin. Nucl. Med. 2023, 53, 426–448. [Google Scholar] [CrossRef]
  51. Urso, L.; Manco, L.; Cittanti, C.; Adamantiadis, S.; Szilagyi, K.E.; Scribano, G.; Mindicini, N.; Carnevale, A.; Bartolomei, M.; Giganti, M. 18F-FDG PET/CT Radiomic Analysis and Artificial Intelligence to Predict Pathological Complete Response after Neoadjuvant Chemotherapy in Breast Cancer Patients. Radiol. Med. 2025, 130, 543–554. [Google Scholar] [CrossRef] [PubMed]
  52. Orlhac, F.; Eertink, J.J.; Cottereau, A.-S.; Zijlstra, J.M.; Thieblemont, C.; Meignan, M.; Boellaard, R.; Buvat, I. A Guide to ComBat Harmonization of Imaging Biomarkers in Multicenter Studies. J. Nucl. Med. 2022, 63, 172–179. [Google Scholar] [CrossRef] [PubMed]
  53. Genta, S.; Ghilardi, G.; Cascione, L.; Juskevicius, D.; Tzankov, A.; Schär, S.; Milan, L.; Pirosa, M.C.; Esposito, F.; Ruberto, T.; et al. Integration of Baseline Metabolic Parameters and Mutational Profiles Predicts Long-Term Response to First-Line Therapy in DLBCL Patients: A Post Hoc Analysis of the SAKK38/07 Study. Cancers 2022, 14, 1018. [Google Scholar] [CrossRef] [PubMed]
  54. Ceriani, L.; Milan, L.; Cascione, L.; Gritti, G.; Dalmasso, F.; Esposito, F.; Pirosa, M.C.; Schär, S.; Bruno, A.; Dirnhofer, S.; et al. Generation and Validation of a PET Radiomics Model That Predicts Survival in Diffuse Large B Cell Lymphoma Treated with R-CHOP14: A SAKK 38/07 Trial Post-hoc Analysis. Hematol. Oncol. 2022, 40, 12–22. [Google Scholar] [CrossRef]
  55. Ortega, C.; Anconina, R.; Joshi, S.; Metser, U.; Prica, A.; Johnson, S.; Liu, Z.A.; Keshavarzi, S.; Veit-Haibach, P. Combination of FDG PET/CT Radiomics and Clinical Parameters for Outcome Prediction in Patients with Non-Hodgkin’s Lymphoma. Nucl. Med. Commun. 2024, 45, 1039–1046. [Google Scholar] [CrossRef]
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 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.
Cancers 17 01827 g001
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 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).
Cancers 17 01827 g002
Figure 3. ROC curves of the CT (a) and PET (b) models with both learners (SVM and RF) in internal validation.
Figure 3. ROC curves of the CT (a) and PET (b) models with both learners (SVM and RF) in internal validation.
Cancers 17 01827 g003
Figure 4. ROC curves of both internal and external validation of CT (a) and PET (b) models.
Figure 4. ROC curves of both internal and external validation of CT (a) and PET (b) models.
Cancers 17 01827 g004
Figure 5. ROC curve comparison (left) and radial plot of the performance scores (right) of SVM learners for CT and PET models.
Figure 5. ROC curve comparison (left) and radial plot of the performance scores (right) of SVM learners for CT and PET models.
Cancers 17 01827 g005
Table 1. Characteristics of the study population in the training and validation cohorts.
Table 1. Characteristics of the study population in the training and validation cohorts.
CharacteristicsTraining Cohort (Rome)
N (%)
Validation Cohort (Ferrara)
N (%)
Patients299
Male sex9 (31%)4 (44.4%)
Median age years (range)35(18–58)39.5(25–58)
IPI score
111 (37.9%)7 (77.8%)
210 (34.5%)1 (0.1%)
33 (10.3%)1 (0.1%)
41 (3.4%)0 (0%)
>40 (0%)0 (0%)
B symptoms9 (31%)0 (0%)
LDH
≤225 U/L19 (65.5%)0 (0%)
>225 U/L10 (34.5%)9 (100%)
ECOG
0–119 (65.5%)9 (100%)
≥210 (34.5%)0 (0%)
Bulky presentation25 (86.2%)8 (88.9%)
Extranodal localization3 (10.3%)0 (0%)
R-based regimen
R-CHOP4 (13.8%)9 (100%)
R-MACOP-B22 (75.9%)0 (0%)
R-MACOP-B + autoTMO2 (6.9%)0 (0%)
R-MACOP-B + surgery1 (3.4%)0 (0%)
EoT PET response
Responders (DS 1–3)23 (79.3%)4 (44.4%)
Not-responders (DS 4–5)6 (20.7%)5 (55.6%)
Consolidative RT
Yes18 (62.1%)8 (88.9%)
No11 (37.9%)1 (11.1%)
Relapse
Yes2 (6.9%)0 (0%)
Not27 (93.1%)9 (100%)
Patient status at last follow-up
Alive with NED28 (96.6%)9 (100%)
Progressive1 (3.4%) 0 (0%)
Median follow-up (years)4.663.64
CHOP = chemotherapy; DS = Deauville score; ECOG = Eastern Cooperative Oncology Group; EoT = end of treatment; LDH = lactate dehydrogenase; MACOP = methotrexate Adriamycin cyclophosphamide oncovin prednisone; NED = no evidence of disease; RT = radiation therapy.
Table 2. Performance scores in 10FoldCV for each ML model.
Table 2. Performance scores in 10FoldCV for each ML model.
Training 10FoldCVAUCCAPRESENSPETPTN
CT ModelRF0.920.880.880.880.8888.90%86.70%
SVM0.970.880.880.880.8884.60%75.00%
PET ModelRF0.830.710.730.710.7177.30%90.90%
SVM0.850.860.850.840.8683.30%61.90%
AUC = area under the curve; CA = classification accuracy; PRE = precision; RF = random forest; SEN = sensitivity; SPE = specificity; SVM = support vector machine; TP = true positive; TN = true negative.
Table 3. Performance scores in the validation step for each ML model.
Table 3. Performance scores in the validation step for each ML model.
Internal ValidationAUCCAPRESENSPETPTN
CT ModelRF0.720.640.650.640.6488.90%66.70%
SVM0.900.930.940.930.9384.60%75.00%
PET ModelRF0.830.710.730.710.7177.30%80.00%
SVM0.860.850.850.840.8583.30%85.70%
External ValidationAUCCAPRESENSPETPTN
CT ModelRF0.850.690.600.600.7560.00%75.00%
SVM0.750.850.970.600.9875.00%77.80%
PET ModelRF0.740.690.570.800.6357.10%83.30%
SVM0.810.770.670.800.7566.70%85.70%
AUC = area under the curve; CA = classification accuracy; PRE = precision; RF = random forest; SEN = sensitivity; SPE = specificity; SVM = support vector machine; TP = true positive; TN = true negative.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Esposito, F.; Manco, L.; Urso, L.; Adamantiadis, S.; Scribano, G.; De Marchi, L.; Venditti, A.; Postorino, M.; Urbano, N.; Gafà, R.; et al. 18F-FDG PET/CT Radiomics for Predicting Therapy Response in Primary Mediastinal B-Cell Lymphoma: A Bi-Centric Pilot Study. Cancers 2025, 17, 1827. https://doi.org/10.3390/cancers17111827

AMA Style

Esposito F, Manco L, Urso L, Adamantiadis S, Scribano G, De Marchi L, Venditti A, Postorino M, Urbano N, Gafà R, et al. 18F-FDG PET/CT Radiomics for Predicting Therapy Response in Primary Mediastinal B-Cell Lymphoma: A Bi-Centric Pilot Study. Cancers. 2025; 17(11):1827. https://doi.org/10.3390/cancers17111827

Chicago/Turabian Style

Esposito, Fabiana, Luigi Manco, Luca Urso, Sara Adamantiadis, Giovanni Scribano, Lucrezia De Marchi, Adriano Venditti, Massimiliano Postorino, Nicoletta Urbano, Roberta Gafà, and et al. 2025. "18F-FDG PET/CT Radiomics for Predicting Therapy Response in Primary Mediastinal B-Cell Lymphoma: A Bi-Centric Pilot Study" Cancers 17, no. 11: 1827. https://doi.org/10.3390/cancers17111827

APA Style

Esposito, F., Manco, L., Urso, L., Adamantiadis, S., Scribano, G., De Marchi, L., Venditti, A., Postorino, M., Urbano, N., Gafà, R., Cuneo, A., Chiaravalloti, A., Bartolomei, M., & Filippi, L. (2025). 18F-FDG PET/CT Radiomics for Predicting Therapy Response in Primary Mediastinal B-Cell Lymphoma: A Bi-Centric Pilot Study. Cancers, 17(11), 1827. https://doi.org/10.3390/cancers17111827

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop