PET‐based radiomics signature can predict durable responses to CAR T‐cell therapy in patients with large B‐cell lymphoma

Abstract Chimeric antigen receptor (CAR) T‐cell therapy is a promising treatment option for relapsed or refractory (R/R) large B‐cell lymphoma (LBCL). However, only a subset of patients will present long‐term benefit. In this study, we explored the potential of PET‐based radiomics to predict treatment outcomes with the aim of improving patient selection for CAR T‐cell therapy. We conducted a single‐center study including 93 consecutive R/R LBCL patients who received a CAR T‐cell infusion from 2018 to 2021, split in training set (73 patients) and test set (20 patients). Radiomics features were extracted from baseline PET scans and clinical benefit was defined based on median progression‐free survival (PFS). Cox regression models including the radiomics signature, conventional PET biomarkers and clinical variables were performed for most relevant outcomes. A radiomics signature including 4 PET‐based parameters achieved an AUC = 0.73 for predicting clinical benefit in the test set, outperforming the predictive value of conventional PET biomarkers (total metabolic tumor volume [TMTV]: AUC = 0.66 and maximum standardized uptake value [SUVmax]: AUC = 0.59). A high radiomics score was also associated with longer PFS and OS in the multivariable analysis. In conclusion, the PET‐based radiomics signature predicted efficacy of CAR T‐cell therapy and outperformed conventional PET biomarkers in our cohort of LBCL patients.


INTRODUCTION
Diffuse large B-cell lymphoma (DLBCL) is the most common type of non-Hodgkin lymphoma [1].Despite durable responses to first-line immunochemotherapy in approximately 60% of cases [2,3], less than half of relapsed/refractory (R/R) patients are salvaged with intensive chemotherapy followed by autologous stem cell transplant consolidation [4].Chimeric antigen receptor (CAR) T-cell therapy is the established third-line treatment for this disease [5,6] and has shown significantly improved results in comparison with chemotherapy for early relapsed and refractory patients in the second-line setting [7,8].
However, not all patients experience the same clinical benefit with this cellular therapy, raising the need to find predictive biomarkers to aid patient selection, leading toward a better efficacy and safety profile.
Quantitative imaging parameters from 18 F-fluorodeoxyglucose (FDG) positron emission tomography (PET) such as total metabolic tumor volume (TMTV) and the maximum standardized uptake value (SUV max ) have been associated with outcomes in patients with DLBCL [9][10][11][12].However, these standard quantitative PET characteristics are limited to tumor burden and maximum metabolic consumption, but do not provide details about the phenotype or spatial distribution of the malignancy.Also, the reported results of these parameters are heterogeneous, and their routine clinical use is still limited [9,11,[13][14][15].
PET-based radiomics provides quantitative imaging data regarding tumor shape, heterogeneity, and metabolic consumption distribution.
These parameters can provide insight into the tumor biology, including meaningful information about the cancer cell metabolism and tumor microenvironment for further identification of responsive phenotypes.
Some studies have defined PET-based radiomic features associated with a favorable response to standard chemotherapy in patients with DLBCL [16,17].However, scarce data is available in the specific setting of CAR T-cell therapy [18].
In this study, we analyzed FDG-PET-based radiomic phenotypes in R/R DLBCL patients treated with CD19-targeted CAR T-cell ther-apy, aiming to identify those patients who would benefit most from treatment.Furthermore, we looked at the added value of textural features to standard prognostic factors for efficacy and safety after therapy.

Study population
The study included all consecutive patients with R/R diffuse large Bcell lymphoma (DLBCL), de novo or transformed from an indolent lymphoma, who received a single infusion of CD19-targeted second-

Endpoints and definitions
Clinical data were collected from the electronic medical records.These  Abbreviations: RAD, radiomics score; SUV max , maximum standard uptake value; TMTB, total metabolic tumor burden.
CAR T-cell infusion and disease progression or death, whichever occurred first.Secondary endpoints included overall survival (OS), defined as the time between CAR T-cell infusion and death, best response achieved after infusion and clinically significant adverse events, defined as grade ≥2 CRS and/or ICANS in this study.

Disease assessment and evaluation
All patients underwent a PET-CT scan at our institution after the last bridging regimen, within 7 days of starting lymphodepleting

PET/CT exam and feature acquisition
The Standard Uptake Value (SUV) was computed voxel-wise from 18 F-FDG PET/CT images normalized to patient's weight and injected dose [22].Lesions were delineated in the PET scans by an experienced nuclear medicine physician using a custom workflow from MIM Encore™ software (MIM Software Inc., Cleveland, OH, USA), which selected all metabolic activity above the 41% maximum SUV threshold.Gray-Level Run Length Matrix [GLRLM], and Gray-Level Size Zone Matrix [GLSZM]).All features were computed as described by van Griethuysen et al. [24].Conventional PET features (SUV max and TMTV) were also computed from the whole metabolic disease.Image preprocessing and feature extraction were computed using Pyradiomics v.3.0.1 [24] package and Python v.3.8.8.

Statistical analysis
Data Cox prediction after discretization by Youden's threshold optimization (Figure 1).The potential impact of radiomic features on safety outcomes (grade ≥2 CRS and/or ICANS) were evaluated using the Mann-Whitney U test.

Patient characteristics
Of 101 consecutive patients with R/R DLBCL treated with CD19targeted CAR T cells during the study period, 7 were excluded because pretreatment PET scans were not available or disease was not measurable, and 1 patient died prior to the first response evaluation (Figure 2).
Finally, 93 patients were included in the study.
Patients' characteristics are summarized in

Radiomics signature development
Seventy-three and 20 patients were randomly selected to build (training set) and validate (test set) the model.Baseline characteristics were similar between the training and test sets (Table 1).Among the 105 computed radiomics features, four of them were identified as those with the highest association with PFS using LASSO with the lambda of λ = 0.102 as described in the methods.The 4 selected features included maximum intensity, skewness, major axis length, and large dependence low gray-level emphasis (LDLGLE) from the texture matrix GLDM.The radiomics scored can be obtained by using Equation ( 1)   2).Accuracy, sensitivity, and specificity for the optimal cut-off of 0.43 are described in Table 2.

F I G U R E 4
Kaplan-Meier curves for the multivariate analysis predicted progression-free survival (PFS) values.

Integration of radiomics and conventional PET features to predict CAR T-cell efficacy
Thereafter, we investigated the performance of conventional PET features to predict PFS in our population.PET established biomarkers including baseline SUV max of all lesions (AUC 0.59 [95% CI 0.32-0.86])and TMTV (AUC 0.66 [95% CI 0.38-0.94]),individually or in combination (AUC 0.69 [95% CI 0.44-0.94]),showed a lower predictive capacity than the radiomics score for 3-month PFS.Furthermore, the addition of preestablished PET parameters to the radiomics score did not increase the predictive capacity of the latter in the training nor test sets (Figure 3).
The low radiomics score and costimulatory domain 4-1BB retained significance in the multivariate LASSO-Cox regression analysis for PFS (Figure 4, Table 3).

DISCUSSION
The  In our study, tumor heterogeneity by radiomics evaluation was associated with a lower probability of response after CAR T-cell therapy.
This could be explained by a diverse composition of the tumor cells and microenvironment and, potentially, heterogeneous expression of the target antigen, which have been described as predictive biomarkers of response in the setting of adoptive T-cell therapies [25,26].
Several limitations should be taken into account in this study.In conclusion, in this study, we trained and validated a proofof-concept PET-based radiomics signature to predict response to CAR T cells.This model outperformed conventional PET parameters and showed an independent predictive value after adjusting for well-known clinical factors.
included age, sex, disease stage before CAR T-cell therapy, previous lines of treatment, previous stem cell transplant (SCT), Eastern Cooperative Oncology Group (ECOG) status, lactate dehydrogenase (LDH), tumor histology (transformed from indolent lymphoma or DLBCL de novo high-grade B-cell lymphoma double hit/triple hit [DH/TH]) and cell of origin (germinal center B-cell [GCB] or non-GCB).The primary endpoint was progression-free survival (PFS), defined as time between F I G U R E 1 Radiomics analysis workflow.F I G U R E 2 ROC curves for 3-month progression-free survival (PFS) prediction of the radiomics signature in the training and test sets (A) and fivefold cross-validation analysis (B).TA B L E 1 Baseline patient characteristics.
chemotherapy.Disease evaluation after CAR T-cell therapy was scheduled at 1, 3, 6, 12, 18, and 24 months after infusion.The imaging reports were based on the Lugano recommendations for response assessment and graded according to the 5-point Deauville score [21].Patients achieving a complete metabolic response (CMR, Deauville scores 1-3) or partial metabolic response (PMR) were considered responders to CAR T-cell therapy.
PET-CT exam required a minimum of 6-hour fast prior to the intravenous administration of 3.7 MBq/kg (222-370 MBq) of 18 fluorodeoxyglucose (18F-FDG).Glucose values below 140 mg/dL were required in all cases prior to administration of the radiopharmaceutical.Before scanning, the patients were at rest for a minimum of 60 min.Images were obtained using a Siemens Biograph mCT, which combines a spiral CT of 64 slices (210 keV, 120 mAs, care dose) with a dedicated PET, from the skull to the upper third of both femurs.The images generated were evaluated by a nuclear medicine physician in a syngo.viaSiemens Healthcare workstation [9].

F I G U R E 3
Area under the reciever operating characteristic (AUROC) curve for 3-month progression-free survival (PFS) prediction of conventional PET biomarkers (SUV max and TMTV) and radiomics score individually or incombination.RAD, radiomics score; SUV, standard uptake value; TMTB, total metabolic tumor burden.Images and segmentations were reviewed by a nuclear medicine specialist.Before feature extraction, images were resampled to 2 × 2 × 2 mm 3 using spline interpolation and discretized to a fixed bin number of 25[23].From the lesion with the largest volume, 105 radiomics features were extracted.These features correspond to quantitative characteristics regarding first order SUV distribution and metabolic consumption heterogeneity in the tumor measured by Gray-Level Cooccurrence Matrix [GLCM], Gray-Level Dependence Matrix [GLDM], was split into training and test sets (80% and 20%, respectively).The training set was used to build the model and the test set to validate its performance.Population characteristics of both sets were balanced according to baseline international prognostic index (IPI) score, costimulatory domain and treatment response.In the training cohort, the least absolute shrinkage and selection operator (LASSO)-regression were used for radiomic feature selection.LASSO regularization parameter (λ) was optimized by maximizing the area under the curve (AUC) from cross-validation.Logistic regression including the selected features was trained and tested for classifying 3-month PFS (median PFS of this cohort).The performance of conventional PET features and the radiomic score were compared using the AUC from the receiver operating characteristic (ROC) curve.The 95% confidence intervals and the p value were assessed with the DeLong method and the Mann-Whitney U test, respectively.Stratified repeated fivefold cross-validation was implemented to explore the model generalizability for different data splits.Finally, univariate and multivariate Cox proportional hazard regression analysis were performed with clinical and PET-based data (including the radiomics signature) to investigate the impact of these factors on continuous PFS and OS.Kaplan-Meier curves were reported for the multivariate ) y = −0.042+ −0.580 × Shape Major Axis Length + −0.331 × First order Maximum + −0.646 × First order Skewness × −0.565 × GLDM LDLGLE.The PET-based radiomics signature points out that having larger lesions (measured by major axis length) negatively contribute to have a worse prognosis.Similarly, lesions with high SUV max (maximum intensity) or negatively skewed distribution (skewness), which means more density of voxels in with high SUV values are more likely to not benefit more from CAR T-cell therapy (

Finally, we
explored the association between the radiomics score and conventional PET-based features on significant adverse events after CAR T-cell therapy.Neither the radiomics score TMTV nor SUV max was associated with a higher incidence of grade ≥2 CRS or ICANS in our cohort (p > 0.05).Similarly, we could not identify risk factors for grade ≥2 CRS and/or ICANS among the clinical variables including age, ECOG ≥1, number of previous lines of treatment, and costimulatory domain (p > 0.05).
advent of CAR T-cell therapy has significantly improved the prognosis of patients with R/R DLBCL.However, only a subset of patients achieves durable responses after therapy.The identification of predictive biomarkers to aid patient selection and prognostic stratification is an unmet need in the field.In this study, we developed a PET-based radiomics signature capable of predicting clinical benefit after CAR T-cell therapy with an accuracy of 75%.This radiomics signature combined metabolic consumption (SUV), tumor size, and metabolic activity distribution, indicating that patients with smaller lesions and lower SUV max or negatively skewed distribution are more likely to benefit from CAR T-cell therapy.The radiomics signature outperformed other established PET parameters, indicating that a more comprehensive evaluation of PET images including radiomic features could, more accuratelycapture, which DLBCL patients are more likely to respond to this novel cellular therapy.Conventional PET-based biomarkers such as TMTV and SUV max have been analyzed as prognostic biomarkers in lymphoma patients treated with CAR T cells [9-11], showing that higher baseline values are associated with a lower probability of response and long-term survival.Moreover, PET-based radiomic signatures have been associated with response to treatment in DLBCL patients treated with conventional chemotherapy, showing a better performance than standard clinical scoring systems, such as the IPI score [16].However, few studies have implemented PET-based radiomics for modeling tumor response to CAR T-cell therapy.Zhou et al. evaluated the prognostic value of different PET-based radiomic features and explored the association between baseline radiomics and cytokine release syndrome Despite being one of the largest studies on radiomics in CAR T-cell patients, our results are limited by the relatively small sample size of the validation cohort.Nevertheless, the results are encouraging to explore further in larger, more heterogeneous datasets, allowing an improvement of the model and additional validation.Of note, our pipeline incorporates early preprocessing steps including SUV standardization, making our model easier to be computed from images acquired with different scanners and facilitating its implementation in clinical practice.Furthermore, we have provided the preprocessing methods as well as the equation to obtain the final score.This will facilitate the validation in of our signature in other external cohorts without needing the data access.Moreover, further analysis of these biomarkers, integrating clinical, imaging and molecular data, could lend insight into the biological mechanisms that lead to higher response rates in these patients.Thus, the indication of CAR T-cell therapy should not be made based on a single variable but integrating radiomics with other prognostic factors.

Disease stage before CAR T-cell therapy
*p Values were obtained from ANOVA test for continuous variables and Fisher's exact test for categorical variables.

Table 3
CI 0.48-0.98] in the training and test sets, respectively, for prediction of 3-month PFS.Moreover, the radiomics signature showed a stable classification performance with an AUC of 0.67 [95% CI 0.48-0.84] in the cross-validation analysis (Figure Univariate and multivariate Cox for progression-free survival (PFS) and overall survival (OS).
TA B L E 3in a small cohort of lymphoma patients.Similar to our results, Zhou et al. found that some textural features had a better performance than conventional PET parameters.Reinert et al. identified temporal associations between radiomic textural features and serologic markers of response such as C-reactive protein, LDH and leukocyte count.Noteworthy, these studies were performed in limited cohorts (less than 30 patients each) and only carried out univariate analysis to associate PET-based radiomic features to outcome.In our study, a multivariate analysis combining different tumor characteristics was performed to define responsive phenotypes.Furthermore, the availability of a larger cohort of patients treated with CAR T cells allowed us to test our model using internal validation.