Prognostic Value of Radiomic Features of 18F-FDG PET/CT in Patients With B-Cell Lymphoma Treated With CD19/CD22 Dual-Targeted Chimeric Antigen Receptor T Cells

Objective In the present study, we aimed to evaluate the prognostic value of PET/CT-derived radiomic features for patients with B-cell lymphoma (BCL), who were treated with CD19/CD22 dual-targeted chimeric antigen receptor (CAR) T cells. Moreover, we explored the relationship between baseline radiomic features and the occurrence probability of cytokine release syndrome (CRS). Methods A total of 24 BCL patients who received 18F-FDG PET/CT before CAR T-cell infusion were enrolled in the present study. Radiomic features from PET and CT images were extracted using LIFEx software, and the least absolute shrinkage and selection operator (LASSO) regression was used to select the most useful predictive features of progression-free survival (PFS) and overall survival (OS). Receiver operating characteristic curves, Cox proportional hazards model, and Kaplan-Meier curves were conducted to assess the potential prognostic value. Results Contrast extracted from neighbourhood grey-level different matrix (NGLDM) was an independent predictor of PFS (HR = 15.16, p = 0.023). MYC and BCL2 double-expressor (DE) was of prognostic significance for PFS (HR = 7.02, p = 0.047) and OS (HR = 10.37, p = 0.041). The combination of NGLDM_ContrastPET and DE yielded three risk groups with zero (n = 7), one (n = 11), or two (n = 6) factors (p < 0.0001 and p = 0.0004, for PFS and OS), respectively. The PFS was 85.7%, 63.6%, and 0%, respectively, and the OS was 100%, 90.9%, and 16.7%, respectively. Moreover, there was no significant association between PET/CT variables and CRS. Conclusions In conclusion, radiomic features extracted from baseline 18F-FDG PET/CT images in combination with genomic factors could predict the survival outcomes of BCL patients receiving CAR T-cell therapy.


INTRODUCTION
Diffuse large B-cell lymphoma (DLBCL) is the most common type of malignant lymphomas in adults, accounting for 30-40% of all non-Hodgkin lymphoma (NHL) worldwide, and DLBCL exhibits a higher prevalence in developing countries (1,2). Indeed, a study from China has found that nearly half of the NHL cases are diagnosed as DLBCL (3). Chimeric antigen receptor (CAR) T-cell therapies targeting CD19 are a promising approach for the treatment of relapsed/refractory (R/R) B-cell malignancies. The complete remission (CR) rate in patients with R/R B-cell acute lymphoblastic leukemia (B-ALL), who received CD19 CAR T-cell therapies, is approximately 90%, while it is only 50% in patients with R/R B-cell NHL (B-NHL) (4)(5)(6). Approximately 30% of relapses after CD19 CAR T-cell therapy are characterized by CD19 antigen loss through a variety of mechanisms, including antigen escape or lineage switch (7,8). Evidence from some studies of solid tumors has shown that compared with single-antigen targeting, dual-or multi-antigen targeting CAR T-cells may result in synergistic effects, which can optimize response rates and prevent antigen escape (9,10). Like CD19, CD22 is expressed in most B-cell malignancies, which is an effective target for CAR T-cell therapy in B-cell hematological malignancies (11).
As a proto-oncogene, MYC plays a central role in the pathogenesis of DLBCL and is particularly involved in its progression (12). The World Health Organization (WHO) classification considers cases with concurrent MYC and BCL2 and/or BCL6 rearrangement determined by fluorescence in situ hybridization (FISH) as "double-hit" lymphomas (DHL) or "triple-hit" lymphomas (THL), respectively. Patients with coexpression of MYC and BCL2 but without underlying rearrangement, as defined by immunohistochemistry (IHC), are commonly referred to as double-expressor (DE) lymphoma (13,14). Studies suggest that patients with co-expression of double proteins have a higher international prognostic index (IPI) score, advanced-stage disease (III/IV), and poor prognoses (15,16). 18 F-FDG PET/CT is already used by clinicians to diagnose and evaluate hematologic malignancies. Recent studies have demonstrated that baseline radiomic features are promising prognostic markers in lymphoma, as they can better predict outcomes compared with conventional imaging metrics (17)(18)(19). Radiomics is the high-throughput extraction of a large number of quantitative image features from medical images and can capture information on the intensity, texture, and shape of lesions. Recently, although radiomics has made significant progress in various malignancies (20,21), radiomic information for R/R B-NHL is still limited.
To the best of our knowledge, no study has introduced textural analysis to predict prognosis in BCL patients treated with CAR T-cell therapies, especially for CD19/CD22-targeting CAR T-cell therapies. Therefore, we aimed to investigate the capacity of radiomic features extracted from baseline 18 F-FDG PET/CT to predict the survival of BCL patients treated with CD19/CD22 dual-targeted CAR T-cell therapies. Moreover, we explored the relationship between baseline radiomic features and the occurrence probability of cytokine release syndrome (CRS).

Patients
This retrospective study was approved by the institutional review board of the First Affiliated Hospital of Soochow University, and the informed consent was waived. This study was carried out following the Declaration of Helsinki with a trial registration number of ChiCTR2100052247.
The inclusion criteria were set as follows: patients were over 18 years old; patients treated with CD19/CD22 dual-targeted CAR T-cell therapies; and patients who underwent 18 F-FDG PET/CT before CAR T-cell infusion. Between June 2017 to July 2021, 24 patients with histologically confirmed BCL were enrolled in the present study. CRS was assessed and graded according to the American Society for Transplantation and Cellular Therapy (ASTCT) criteria (22). Neurotoxicity was assessed and graded according to the Common Terminology Criteria for Adverse Events. Cutoff values of 40% for MYC protein and 50% for BCL2 protein were defined as the DE status (13).

PET/CT Acquisition
The 18 F-FDG PET/CT examination was performed after 6 h of fasting with blood glucose lower than 11.1 mmol/L. Approximately 40-60 min after the injection of 18 F-FDG (4.07-5.55 MBq/kg), PET/CT was performed from the base of the skull to the midthigh with 2-3 min per bed position (reconstructed by ordered subset expectation-maximization algorithm) using a Discovery PET/CT (General Electric Medical Systems, Milwaukee WI, USA) with low-dose CT parameters (140 kV, 120 mA, transaxial FOV of 70 cm, slice thickness 3.75 mm).

Feature Extraction and Selection
Quantitative PET/CT analysis was performed using the LIFEx freeware (v6.30 https://www.lifexsoft.org/) (23). The volume of interest (VOI) was contoured manually on co-registered images by two experienced nuclear medicine physicians who were blinded to the clinical and pathological information of patients. The whole layers in three-dimensional VOI were delineated on each slice, and 41% of the maximum standardized uptake value (SUV max ) was used as the threshold to define VOI (24). The details of the tumor segmentation are described in Figure 1. Spatial resampling had a voxel size of 2 × 2 × 2 mm. Intensity discretization for CT data was performed with the number of gray levels of 400 bins and absolute scale bounds from -1,000 and 3,000 HU, while it was conducted with 64 bins between 0 and 20 for PET data. The radiomic features were extracted from both PET and CT images within the same VOI due to the good matching of PET and CT images. The radiomics workflow is shown in Figure 2. A total of 92 radiomic features were extracted, including 47 PET-derived features and 45 CT-derived features ( Table S1).
The radiomic features of PET and CT images were selected by the following procedure. Intraclass correlation coefficient (ICC > 0.75) was first performed to remove the redundant features (25,26). Subsequently, the retained features were further selected by the least absolute shrinkage and selection operator (LASSO) regression algorithm (27). In addition, 10-fold cross-validation was applied to select the parameter of Lambda (l) to avoid overfitting.

Statistical Analysis
The categorical variables were analyzed using chi-square statistic or Fisher's exact test. Mann-Whitney U-test was applied to continuous variables. Progression-free survival (PFS) and overall survival (OS) were defined as the time from first CAR T-cell infusion until disease progression or death from any cause, respectively. The optimal cutoff values for PFS and OS were determined using receiver operating characteristic (ROC) curves according to the Youden index. Survival functions were calculated by the Kaplan-Meier method, and comparisons between subgroups were made using a log-rank test. Univariate and multivariate analyses were carried out using Cox proportional-hazards models. P < 0.05 was considered statistically significant. IBM SPSS statistics version 26.0, PYTHON version 3.0 (https://www.python.org), and MedCalc software (MedCalc Software, Ostend, Belgium) were used for statistical analyses.

Patient Characteristics
A total of 24 patients (median age of 51 years, range of 26-70 years) who received CAR T-cell therapy were enrolled in the present study. Table 1 summarizes the baseline characteristics of the patients. For the histological type, 20 (83.33%) patients The median OS for the entire group was not reached. OS rates at 1 and 2 years were 79.17% and 79.17%, respectively ( Figure 3D). The median PFS was 13 months. The 1-and 2year PFS rates were 58.33% and 58.33%, respectively ( Figure 3C). Six patients died after infusion, with a median time of 2.5 months (range of 1-37 months), and all deaths were attributed to the progression of lymphoma ( Figures 3A, B). Moreover, 11 patients relapsed/progressed at a median time of 2 months (range of 6 to 37) after infusion.

Feature Selection
A total of 92 radiomic features were extracted from PET and CT images. Based on the LASSO regression model, we obtained five and four radiomic features for PFS and OS, respectively. For PFS, the CT zone percentage from gray-level zone-length matrix (GLZLM_ZP CT ), PET run-length non-uniformity from graylevel run-length matrix (GLRLM_GLNU PET ), Contrast from neighborhood gray-level difference matrix (NGLDM_ Contrast PET ), SHAPE_Volume PET , and SHAPE_Sphericity  (SHAPE_Sphericity PET ) were selected. For OS, CT long-zone high gray-level emphasis from gray-level zone-length matrix (GLZLM_LZHGE CT ), PET Energy from grey-level cooccurrence matrix (GLCM_Energy PET ), NGLDM_Contrast PET , and zone-length non-uniformity from gray-level zone-length matrix (GLZLM_ZLNU PET ) were selected. The ICC of the radiomic features was all above 0.75. Table 2 summarizes the ROC analysis results of PFS and OS.

Characteristics and Outcomes
After a median follow-up of 10.5 months (range of 1-48 months), the overall response (OR) rate, evaluated by 18 F-FDG PET/CT or CT at 1 month after infusion, was 87.5% (21/24), 50% (14/24) patients achieved CR, and 29.17% (7/24) patients achieved partial response (PR). The median NGLDM_Contrast PET was 0.350 (range of 0.000-1.296) in patients who achieved CR, which was significantly lower compared with patients who did not achieve CR (median 0.635 [range of 0.046-1.177]; p= 0.02; Figure 4). Patients who received less than two lines of chemotherapy had a higher probability of CR (p= 0.03). No significant association was found between other characteristics and outcomes in this cohort (all p>0.089, Tables S2 and S3).

Prognostic Model Construction
A clinical model was constructed based on multivariate Cox analysis of significant risk factors. The risk factors included higher NGLDM_Contrast PET (>0.468) and positive DE. The clinical and PET models were combined, and all patients were stratified into three risk categories with distinct PFS (p < 0.0001) and OS (p = 0.0004): group I (no risk factors; n = 7); group II (one risk factor only; n = 11); and group III (two risk factors; n = 6). These three groups had significantly different PFS, which was   Figure 5F).

Baseline Variables and Adverse Effects
In our present study, 14 patients (58%) experienced CRS, including seven (29%) patients with Grade 1 CRS, three (13%) patients with Grade 2 CRS, and four (16%) patients with Grade 3 CRS. All patients responded to established conventional therapies, and no patients received tocilizumab treatment. The PET/CT variables in patients with non-severe CRS (0-2) were not significantly different from those who had severe CRS (3-4) (all p > 0.05, Figure 4 and Table S4). No patient developed neurotoxicity.

DISCUSSION
In the current study, we investigated the potential prognostic value of radiomic features derived from 18 F-FDG PET/CT images in BCL patients treated with CD19/CD20-targeting CAR T-cell therapies. The results showed that DE and NGLDM_Contrast PET were two independent prognostic factors for PFS, whereas DE and MYC expression were of prognostic significance for OS. Moreover, a prognostic stratification model was established to identify risk stratification of patients by integrating clinicopathological and PET/CT imaging prognostic factors. Our findings indicated that the baseline PET/CT-based radiomic features might contribute to the risk stratification of BCL patients. Radiomics is an emerging method to acquire a large amount of high-dimensional data through the analysis of entire disease lesions and can quantify the non-uniformity between adjacent voxels. Measuring the textural indices of PET/CT images to quantify intratumor heterogeneity has been proposed as an adjunct to predict outcomes (17,18,20). Moreover, due to the differences in radiomic features and numbers extracted, it is difficult to make direct comparisons between studies. Lue et al. (17) have found that baseline PET-radiomic feature (GLRLM_RLN) can be used to predict the survival outcomes of DLBCL. Cheng et al. (28) have suggested that uniformity from NGLCM is a significant prognostic factor for patients with oropharyngeal squamous cell carcinoma. To the best of our knowledge, this was the first report to investigate the potential of FDG PET/CT radiomic features for outcome prediction in BCL patients treated with CAR T-cell therapies. To reduce the impact of discretization values on robustness, an absolute resampling was used. Several radiomic features extracted from PET and CT images were prognostic factors: including GLZLM_ZP CT , GLRLM_GLNU PET , NGLDM_Contrast PET , SHAPE_Volume PET ,  SHAPE_Sphericity PET , GLZLM_LZHGE CT , GLCM_Energy PET , and GLZLM_ZLNU PET . Finally, the multivariable analysis revealed that Contrast extracted from NGLDM appeared to be the independent predictor of PFS. NGLDM is expressed as greylevel variability between one voxel and its neighbors in three dimensions. Higher NGLDM_Contrast PET was associated with poor prognosis, indicating the importance of tumor heterogeneity on PET as a predictor of disease progression.
Genetic rearrangements of MYC and BCL2 and lymphoma derived from MYC and BCL2 co-expression have received increasing attention in recent years (29). DE lymphoma accounts for 18% to 42% of DLBCL cases, which may be related to gene amplification or translocation (29)(30)(31). In our present study, 45.8% of the analyzed patients demonstrated coexpression of MYC and BCL2 proteins. Similarly, in a study by Xu et al. (31), 42% of the patients show these aberrations.  Besides, co-expression of MYC and BCL2 is a prognostic biomarker in the management of DLBCL (32). Recent data have demonstrated that patients with DE are more likely to have high-risk gene signatures and more inferior prognoses (33,34). These findings emphasize the value of DE as a promising tool for prognostic predictors. Compared with fluorescence in situ hybridization (FISH), co-expression of MYC and BCL2 defined by IHC assay represents an inexpensive, rapid, and reproducible technique that has broader application prospects in clinical practice.
In the present study, NGLDM_Contrast PET and DE had independent predictive values. We suggested establishing a prognostic scoring system based on these two characteristics that were complementary in that sense, and they were characterized by two different aspects of the disease: radiomic features and genomic features. This score appeared to have a higher capability for patient risk stratification. Combination of NGLDM_Contrast PET and DE stratified the population into three different prognostic groups; group I (low NGLDM_Contrast PET and negative-DE; PFS 85.7%, OS 100%), group II (high NGLDM_Contrast PET or positive-DE; PFS 63.6%, OS 90.9%), and group III (high NGLDM_Contrast PET and positive-DE; PFS 0%, OS 16.7%). Our results suggested that the combination of these two parameters could provide promising prognostic information for B-NHL patients.
Few studies have investigated the role of MTV in predicting treatment outcomes in BCL (35)(36)(37). In lymphoma, several reports have also indicated that PET/CT radiomic features are significantly associated with survival, whereas conventional PET metabolic parameters (SUV, MTV, and TLG) are not (17,38). In another study with NHL patients treated with CD19-targeting CAR T-cell therapy, TMTV does not have a significant association with OS (39). In contrast, Wang et al. (40) have reported that traditional imaging parameters are more efficient than textural features for predicting therapeutic response and survival. Dean et al. (41) have reported that baseline TMTV, in large B-cell lymphoma (LBCL) patients who are treated with axicabtagene ciloleucel (axi-cel), has a significant prognostic impact on PFS and OS. In our present study, MTV was a prognosticator of both PFS and OS in the univariate analysis. However, we noticed that the statistical significance was not achieved in the multivariate analysis, which was presumably attributed to the small sample size. Many factors may contribute to these differences, such as small sample size, reconstruction parameters, segmentation, and software (25). Our results indicated that the features of intratumor heterogeneity might serve as a complementary indicator and outperform MTV. Further studies are required in a larger cohort population to validate our findings.
In agreement with the previous studies (42), we did not observe any association between PET/CT parameters and CRS, possibly because of the relatively limited number of patients included and the small number of high-grade CRS (≥2) in patients receiving CD19/CD22 CAR T-cell therapies.
There are several limitations to our study. First, it lacked a patient cohort for external validation, and these results should be validated in a large cohort of patients. Second, the number of patients included in this study was relatively small due to the novelty of the therapy. Other limitations of the study derived from the single-center and retrospective data.

CONCLUSIONS
In conclusion, radiomic features extracted from baseline 18 F-FDG PET/CT images in combination with genomic factors could predict the survival outcomes of BCL patients receiving CAR Tcell therapy. No significant association was found between PET/ CT parameters and CRS. Further prospective studies and validation in a large cohort of patients are required to confirm these results.

DATA AVAILABILITY STATEMENT
The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary Material.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by the First Affiliated Hospital of Soochow University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

AUTHOR CONTRIBUTIONS
YZ, JL, XZ, and SD conceptualized and designed the study. TJ, BZ, ND, and SS performed analysis. SS and SD interpreted the data. YZ and JL drafted the manuscript. SS and SD revised the manuscript. All authors contributed to the article and approved the submitted version.