Development of a radiomic-clinical nomogram for prediction of survival in patients with diffuse large B-cell lymphoma treated with chimeric antigen receptor T cells

In our current work, an 18F-FDG PET/CT radiomics-based model was developed to assess the progression-free survival (PFS) and overall survival (OS) of patients with relapsed or refractory (R/R) diffuse large B-cell lymphoma (DLBCL) who received chimeric antigen receptor (CAR)-T cell therapy. A total of 61 DLBCL cases receiving 18F-FDG PET/CT before CAR-T cell infusion were included in the current analysis, and these patients were randomly assigned to a training cohort (n = 42) and a validation cohort (n = 19). Radiomic features from PET and CT images were obtained using LIFEx software, and radiomics signatures (R-signatures) were then constructed by choosing the optimal parameters according to their PFS and OS. Subsequently, the radiomics model and clinical model were constructed and validated. The radiomics model that integrated R-signatures and clinical risk factors showed superior prognostic performance compared with the clinical models in terms of both PFS (C-index: 0.710 vs. 0.716; AUC: 0.776 vs. 0.712) and OS (C-index: 0.780 vs. 0.762; AUC: 0.828 vs. 0.728). For validation, the C-index of the two approaches was 0.640 vs. 0.619 and 0.676 vs. 0.699 for predicting PFS and OS, respectively. Moreover, the AUC was 0.886 vs. 0.635 and 0.778 vs. 0.705, respectively. The calibration curves indicated good agreement, and the decision curve analysis suggested that the net benefit of radiomics models was higher than that of clinical models. PET/CT-derived R-signature could be a potential prognostic biomarker for R/R DLBCL patients undergoing CAR-T cell therapy. Moreover, the risk stratification could be further enhanced when the PET/CT-derived R-signature was combined with clinical factors.


Introduction
Chimeric antigen receptor T cell (CAR-T) therapies are widely used in treating aggressive relapsed or refractory (R/R) B cell malignancies, such as diffuse large B-cell lymphoma (DLBCL), with complete remission rates of about 43% Schuster et al. 2017;Breen et al. 2020). However, many patients experience disease progression after CAR-T therapy. Thus, it is critical to enhance risk stratification for monitoring clinical response and tailoring treatment regimens.
Quantitative 18 F-fluorodeoxyglucose positron emission tomography-computed tomography ( 18 F-FDG PET/CT) parameters can predict outcomes in DLBCL, such as maximum standardized uptake value (SUV max ), metabolic tumor Yeye Zhou, Bin Zhang, and Jiangqin Han contributed equally to this work.
1 3 volume (MTV), and total lesion glycolysis (TLG) (Schmitz et al. 2020;Dean et al. 2020;Song et al. 2016;Cottereau et al. 2016). However, none of the above-mentioned factors can accurately embody the spatial distribution of 18 F-FDG, which is associated with tumor heterogeneity and poor outcomes (Yan et al.2015). Radiomics is a new and promising field, and it can extract many quantitative imaging characteristics from diagnostic images to quantify intratumoral metabolic heterogeneity (Moon et al. 2019;Choi et al. 2016;Dissaux et al. 2020). It has excellent diagnostic and prognostic values for some solid tumors (Li et al. 2021;Du et al. 2021;Lee et al. 2021). However, few studies have reported the prognostic significance of PET/CT radiomic features in DLBCL patients receiving CAR-T therapy (Zhou et al. 2022).
Therefore, a radiomics model was developed to predict overall survival (OS) and progression-free survival (PFS) in DLBCL cases receiving CAR T-cell therapy.

Patients
This work was authorized by the local ethics committee of the First Affiliated Hospital of Soochow University, and no written consent was required because this was a retrospective study (ChiCTR2100052247).
The medical records and imaging data of DLBCL patients undergoing CAR-T cell therapy between March 2017 to Jan 2022 were retrospectively reviewed. Inclusion criteria were as follows (1) ≥ 18 years old; (2) pathologically confirmed R/R DLBCL; (3) 18 F-FDG PET/CT performed before CAR-T cell therapy; (4) availability of clinical-pathologic data; and (5) patients must have evidence of measurable disease. In addition, cases were ruled out if they had incomplete clinical or imaging datasets or other types of malignancies or if patients were treated with granulocyte colony-stimulating factor (G-CSF) within 1 month prior to PET/CT scan. A total of 61 patients (mean age of 52.08 ± 13.00 years) were enrolled in the study and randomly assigned into two cohorts at a ratio of 7:3 (Zhao et al. 2022). These studies involved 61 patients, including PET/CT scans performed at diagnosis in eight patients and PET/CT scans conducted straightway before the infusion of CAR-T therapy in 53 patients. The American Society for Transplantation and Cellular Therapy (ASTCT) criteria were adopted to grade cytokine release syndrome (CRS) and neurotoxicity (Lee et al. 2018).

PET/CT imaging
18 F-FDG PET/CT examinations were performed on a Discovery PET/CT (General Electric Medical Systems, Milwaukee, WI, USA) with standard CT indexes (140 kV, 120 mA, transaxial FOV of 70 cm, slice thickness of 3.75 mm). The level of blood glucose was less than 11 mmol/L before the 18 F-FDG injection. After no less than 6 h of fasting, 18 F-FDG (4.07-5.55 MBq/kg) was intravenously injected 40-60 min prior to image acquisition. Whole-body PET/CT data were acquired according to our standard protocol.

Segmentation and feature extraction
18 F-FDG PET/CT images were analyzed using the LIFEx freeware (version 6.30 https:// www. lifex soft. org/) (Nioche et al. 2018) by two senior physicians with knowledge of nuclear medicine. Lesions were defined as areas with abnormal uptake of 18 F-FDG on PET and abnormal density on CT. The volume of interest (VOI) in three-dimensional coordinates was automatically defined on PET images with the SUV max threshold of 41% (Boellaard et al. 2015), including all nodal and extranodal lesions. The radiomics features are summarized across all VOIs. Because the PET and CT images matched well, the radiomics features of the CT and PET images could be extracted from the same VOI. When the PET and CT images did not match, the VOI for PET and CT were drawn separately. Spatial resampling with 2-mm spacing in X, Y, and Z was conducted on PET and CT images. Intensity discretization was automatically conducted based on gray levels of 64 bins and absolute scale bounds between 0 and 20 for PET data, while that of CT data was done with 400 bins between -1,000 and 3,000 HU (Ou et al. 2020). Finally, 92 radiomic features were isolated, including 45 traits from the CT dataset and 47 from the PET dataset. All the radiomic traits were given in the Supplemental Materials.

Radiomic feature selection and model building
First, the radiomic features with high interobserver repeatability (ICC > 0.75) were selected from the training cohort (Zwanenburg et al. 2019;Jha et al. 2021). Next, the most valuable prognostic features were selected using the least absolute shrinkage and selection operator (LASSO) method. Four PET radiomic features (GLCM_Entropy_log10, HISTO_Entropy_log10, HISTO_Skewness, and GLZLM_ LZE) and two CT radiomic features (SHAPE_Sphericity and GLZLM_SZHGE) were selected for predicting OS. Four PET radiomic features (GLRLM_SRLGE, HISTO_Skewness, GLCM_Entropy_log10, and GLZLM_SZE) and two CT radiomic features (NGLDM_Contrast and SHAPE_ Compacity) were chosen for assessing PFS. The radiomic signatures (R-signatures) were computed using the linear sum of the chosen radiomic features with their respective coefficients weighted (Tibshirani et al. 1997).

3
Finally, the receiver operating characteristic (ROC) curves were used to determine the optimal threshold for R-signatures, and patients were divided into high-risk and low-risk groups .
The significant prognostic parameters of PFS and OS in the training cohort were chosen by univariable and multivariable Cox regression analyses, respectively. All the significant PET/CT and clinical factors in the univariate analysis were input into a multivariate Cox regression. Independent predictors were adopted to construct the radiomics model. Likewise, all important clinical indexes were input into a multivariate Cox regression to establish the clinical model. Radiomic and clinical nomograms were provided.

Model validation
Harrell's concordance index (C-index) and ROC curve were used to evaluate the predictive performance of the models in the training and validation cohorts. Calibration curves were generated to evaluate the calibration of the models. Decision curve analysis (DCA) was used to analyze the clinical usefulness of different models.

Statistical analysis
Statistical analyses were conducted with SPSS version 26.0 and R software 4.2.0 (http:// www.R-proje ct. org). Chisquared test or Fisher's exact test was adopted to compare categorical variables between the training and validation cohorts. PFS and OS were defined as the time from the first CAR T-cell infusion until disease progression or death from any cause, respectively. ROC curves were used to determine the optimal threshold for R-signatures, and patients were divided into high-risk and low-risk groups. Univariate and multivariate Cox regression analyses were adopted o assess the prognostic parameters for PFS and OS. Survival curves were estimated by the Kaplan-Meier method, and the two subgroups were compared using the log-rank test. The calibration curves, C-index, and DCA were calculated for the models in the training and validation cohorts. The threshold of significance was set at P < 0.05. Table 1 exhibits the patients' clinical information in the training and validation cohorts. No significant difference was detected between the two cohorts (P = 0.370-0.822). In addition, follow-up data indicated that the mortality rate was 59.5% in the training cohort, and tumor relapse was found in 40.5% of patients, which was not significantly different from the validation cohort (57.9% and 42.1%, respectively) (P > 0.05).

Radiomics feature selection and R-signature construction
Prognosis-related traits were chosen from the LASSO regression in the training cohort. As a result, four PET radiomic features and two CT radiomic features were chosen for assessing OS and PFS. Besides, the R-signature was determined for assessing OS (AUC = 0.781) and PFS (AUC = 0.768) according to the above-selected radiomics features. The optimal cut-off values of the R-signatures were 0.405 for OS and 0.624 for PFS in the training cohort. Accordingly, patients were classified into the low-risk and high-risk groups, and 61.90% (26/42) and 23.81% (10/42) of patients were classified as "high-risk" in OS and PFS, respectively. Figure 1 presents the radiomic workflow.

Model construction
The relationship between clinical features and imaging indicators with survival was evaluated using univariate and multivariate Cox regression analyses (Table 2  and Table 3 TLG (HR = 2.378, P = 0.043), and R-signature PFS (HR = 3.128, P = 0.014) were associated with PFS. Higher R-signatures showed remarkable correlations with worse PFS and OS in the training and validation cohorts (Fig. 3). Figure 4 shows a PET/CT case of a high-risk patient with a progressively shorter PFS. Moreover, the ECOG and R-signature OS were statistically significant in assessing OS and were adopted to construct the radiomics model. Similarly, the radiomics model for PFS prediction was established using R-signature PFS and Grade of CRS. At the same time, we constructed a clinical model after we analyzed the clinical variables in the training cohort using multivariate analysis. The clinical model was established for OS by the Grade of CRS and DE, and the Grade of CRS and ECOG were integrated for PFS.

Assessment and validation of the models for assessing PFS and OS
To assess patients 1-, 2-and 3-year PFS and OS, we developed a radiomics nomogram and a clinical nomogram (Fig. 5).
For the training cohort, the C-indices and the AUC of the radiomic model were 0.780 (95% CI 0.663-0.897) and 0.828 (95% CI 0.693-0.963) for OS prediction, respectively. The radiomic model performed slightly better than the clinical model (C-index: 0.762, 95% CI 0.650-0.874, and AUC: 0.728, 95% CI 0.567-0.890). For PFS prediction, the C-indices and the AUC of the radiomic model were 0.710 (95% CI 0.607-0.814) and 0.776 (95% CI 0.635-0.918), respectively. The radiomic model performed slightly better than the clinical model (C-index:    (Table 4).  The calibration curves of the nomograms showed good agreement between the estimated and actual observations in the training set of the radiomics model (Fig. 5). Figure 6 shows the decision curve analyses for both models. Again, DCA displayed that the radiomics nomogram had a greater net benefit than the clinical nomogram for both PFS and OS.

Discussion
The potential prognostic value of radiomic traits derived from pre-infusion 18 F-FDG PET/CT images in refractory/ relapsed DLBCL patients receiving CAR T-cell therapies was investigated. Our data indicated that the radiomics model integrating clinical features and R-signatures could be a potential tool to identify individual characteristics to guide personalized treatment. Since SUV-based methods are affected by several physiological and technical factors, including patient preparation, harmonization of image acquisition, reconstruction, and analysis (Schöder at al. 2016;Zaucha at al. 2019), other features derived from PET/CT images are currently being explored. In radiomics, advanced computational methods are applied to medical imaging data to transform medical images into quantitative tumor tissue descriptors (Aerts at al. 2014). The first experience with radiomics in malignant lymphomas was reported by Ben Bouallegue et al. (Ben Bouallègue at al. 2017), showing that most features and shape analyses extracted from baseline 18 FDG PET/CT images are considered to be significantly associated with complete metabolic response, in particular MTV, TLG, surface extension, 2D and 3D fractal dimensions. Given this background, we first attempted to develop a radiomics-based model by combining PET/CT-based radiomic features with clinical features to investigate the potential prognostic value of DLBCL patients receiving CAR T-cell therapy. In our present study, cox regression analysis demonstrated that R-signature on pre-infusion 18 F-FDG PET/CT was an independent predictor of both OS and PFS. Moreover, radiomic-based and clinical models were constructed to assess disease progression. The radiomics nomogram showed superior predictive performance than the clinical nomogram in our training and validation cohorts, indicating that R-signature composed of multiple radiomic features could provide more prognostic information than clinical factors to better identify patients with different outcomes. Patients with higher R-signature had poor survival. These data supported previous findings, indicating that radiomic features derived from medical images could affect the evaluation of disease status and risk stratification in DLBCL (Lue at al. 2020;Aide at al. 2020;Eertink at al. 2022). Kostakoglu et al. (Kostakoglu at al. 2022) have assessed the potential prognostic value of radiomic features derived from PET images in patients with previously untreated DLBCL. The random forest model with COO subgroups identified a clearer high-risk population than IPI. This finding suggested that  PET-based radiomic features combined with well-known risk factors improved the prognostication of patients with untreated DLBCL compared with traditional clinical risk factors alone. Our data indicated that traditional PET factors (SUV max , MTV, and TLG) were not selected for building the models. In contrast, these metabolic factors are significant prognostic biomarkers for DLBCL (Schmitz at al. 2020;Dean at al. 2020;Song at al. 2016;Iacoboni at al. 2021). Such a discrepancy might be attributed to the different feature selection and model construction methods. Our present work used the LASSO regression algorithm to screen optimal features most strongly linked to patient survival and avoid overfitting (Decazes at al. 2018;Ceriani at al. 2022;McEligot at al. 2020). Only radiomic features were ultimately chosen by this algorithm, suggesting that radiomic traits had better discrimination than traditional PET parameters in predicting therapeutic response and survival.
Radiomics analysis could be affected by many factors, including extraction software platform choice, reconstruction factors, and segmentation factors (Barrington at al. 2019;Fornacon-Wood at al. 2020). In our present study, we used a 41% threshold segmentation method to define the VOI, and ICC analysis showed that the majority of radiomic features exhibited superior interobserver repeatability (ICC ≥ 0.75), supporting previous findings Mayerhoefer at al. 2019;Lue at al. 2020). Besides, all PET/ CT images in this study were realized in the same center using the same acquisition and reconstruction protocols.
This study has some limitations. First, although we had the largest patient cohort compared with other investigations, the sample size was still inadequate for accurate analysis. Second, only the dataset from a single institution was adopted to establish and verify the models. Future investigations with multicenter prospective studies in a larger cohort are necessary to verify our results.

Conclusions
In conclusion, prognostic models combining R-signature and clinical risk factors had the potential to assess therapeutic efficacy in patients with R/R DLBCL receiving CAR-T cell therapy.