Non-Hodgkin’s lymphoma (NHL) accounts for approximately 4% of all cancers, with over 20,250 deaths in 2021 in the United States (1–3). Diffuse large B cell lymphoma (DLBCL) is the most common form of aggressive NHL, accounting for 30% to 50% of cases (4, 5). DLBCL is genetically and biologically heterogeneous, with variable therapy responses (6–8). Axicabtagene ciloleucel (axi-cel), a CD19-targeted chimeric antigen receptor (CAR)-T cell therapy, has revolutionized the treatment of the disease, demonstrating superior levels of durable response in relapsed/refractory (R/R) DLBCL (9–11). As only a proportion of patients respond to this relatively new therapy, developing prognostic biomarkers is a critical unmet clinical need.

Total metabolic tumor volume (MTV) measured on ¹⁸F fluorodeoxyglucose (¹⁸F-FDG) positron emission tomography/computed tomography (PET/CT) has been associated with treatment outcomes, and the assessment involves a semi-automated process with clinical oversight to avoid false detection (12, 13). Recent advancements in imaging methodologies have made it possible to utilize medical images (PET/CT) to develop quantitative biomarkers using radiomic metrics (14–16), further extended by artificial intelligence (AI) methods (17, 18). Recently, radiomic metrics on PET imaging have been shown to be prognostic for complete response (19) and predict clinical benefit after CAR-T treatment (20).

In this study, we used advanced quantitative imaging metrics (radiomics) on PET/CT images of DLBCL tumors to identify imaging observed metrics that may predict treatment response (1-year survival), specifically to CAR-T therapy. We compared our findings with whole-body MTV. Our study overview is illustrated in Supplementary Figure S1 .

In this study, we created a cohort of 155 patients who were treated with axi-cel therapy. Baseline radiological scans (PET/CT) scans were obtained and there were 1,639 lesions assessed across the patient cohort that our clinical radiologist manually verified. Of these, 1,058 lesions were detected in 100 patients (MCC) and 581 lesions were detected in 55 patients from ZUMA-1 (consortium). The normal (benign) hepatic (liver) reference region was excluded from the counts. We formed sub-cohorts with patients with their largest lesion related to either lymphatic (n=124) or extranodal (n=94) disease. Figure 1 shows lymphatic (i.e., nodal) and extranodal lesions. The most frequently involved nodal sites were the abdomen (46.77%, n=58), pelvis (17.74%, n=22), mediastinum (8.87%, n=11), neck (8.06%, n=10), and lungs (5.65%, n=7). The most frequently involved extranodal sites were the lungs (36.17%, n=34), musculoskeletal system (21.28%, n=20), pelvis (8.51%, n=8), liver (7.45%, n=7), and spleen (4.26%, n=4), ( Table 1B ). The fifth most common extranodal site in our cohort was evenly shared between the spleen, legs, and skin, with approximately 4.26% of patients (n=4). In this study, radiomic features were categorized into major functional characteristics, i.e., Size, Shape, and Texture (n=259 features) ( Supplementary Table S1 ), and feature metrics were assessed independently in CT and PET scans. We reviewed the relationship of principal components across three categories (Size, Shape, Texture) with MTV and individual lesion volume, displayed as a scatter plot of the principal components, PC1 and PC2, for extranodal CT features and extranodal PET features, respectively ( Supplementary Figure S2 ).

The dependency between the radiomic features and whole-body MTV was evaluated by correlating these metrics at the category level, and we found that the Texture feature PC1 (first principal component) in the Lymphatic group (CT features) had the lowest correlation (ρ =0.35), followed by the Texture PC1 feature in the Extranodal group (CT features, ρ =0.38). The details are given in Supplementary Table S4 , Supplementary Figure S2 .

A logistic regression model was built using total body MTV and radiomic feature-based PCs to predict survival (1 year) after treatment. We found that total body MTV remains predictive with an average AUC of 0.71 to 0.76 across the different sub-cohorts (see Tables 2A–D ). We built a logistic model using the Shape-based PCs, which were computed on CT (Lymphatic group) and MTV, either independently or in combination. We found that MTV shows better predictive value than the Shape-based PCs, with an average AUC of 0.72 and 0.53, respectively, with the combination not showing an improvement (average AUC of 0.67) (see Table 2A ). In the PET SUV (Lymphatic group) image cohort, using Shape-based PC features had an average AUC of 0.61, and the MTV-based predictor yielded an average AUC of 0.71. The combination of these metrics marginally improved the predictive value (average AUC to 0.72), while adding LDH did not improve the predictive ability (AUC of 0.62) (see Table 2B ). Using CT image-based features (Extranodal group) did not show predictive value improvement with the addition of MTV (Shape-based PCs avg. AUC was 0.57 [0.31,0.83], MTV was 0.76 [0.62,0.89], while the combination was 0.46 [0.33,0.60], and with LDH it was 0.697 [0.51, 0.89]) (see Table 2C ). Using PET SUV image metrics (Extranodal group) had an average AUC of 0.68[0.53,0.82], while MTV had an average AUC of 0.74[0.58, 0.89], and the combination showed an improved average AUC of 0.78 [0.7, 0.87]; however, the addition of LDH did not improve the predictive value (avg. AUC of 0.68 [0.42,0.94]) (see Table 2D , Figure 2 ). We further divided the cohort into lower tumor burden (lower median MTV) and higher tumor burden (greater than median MTV). We repeated the predictive analysis to evaluate the role of radiomic metrics. We found that the Shape-based PCs showed a comparable predictive value of AUC 0.716 to MTV (AUC of 0.75), (see Table 2A ). The 1-year OS predictive ability did not improve in combination (Shape-based PCs and MTV). Furthermore, in patients with higher tumor burden, MTV and the Shape-based PC did not seem to be good predictors of 1-year OS (see Table 2B ). We found that the Texture-based PCs showed a similar trend to the Shape-based PCs. The Texture-based PCs computed in CT (Lymphatic group), PET SUV (Lymphatic group), and CT (Extranodal group) did not improve on MTV-based predictors in the respective cohorts (see Supplementary Tables S8A–C ). When using the Texture-based PCs computed on PET SUV (Extranodal group), the metrics had a moderate predictive ability, with an average AUC of 0.59 compared to MTV-based predictors (AUC of 0.74), while the combination showed a moderate improvement (average AUC of 0.66) (see Supplementary Figure S3 , Supplementary Table S8D ). We tested for prognosis using KM plots by grouping patients based on the logistic model and found that the Shape-based PC and Texture-based PC models were significant (P<0.001) for overall and progression-free survival (see Figure 3 ; Supplementary Figure S4 ). Using Shape features on PET images (Non-lymphatic group) had an 18% increased risk of disease progression compared to 15% using MTV, with a CI of 1.018, 1.371 and 1.038, 1.281, respectively (see Supplementary Table S6 ). The patients in our study had a median follow-up of 1 year after CAR-T treatment. Figure 4 shows representative patient scans selected based on Shape-based PC predictors (high and low) for extranodal PET features.

In this study, we developed and validated a quantitative radiological imaging signature derived from DLBCL PET/CT scans that could be used as a biomarker to predict response in patients treated with axi-cel treatment and allow us to identify patients that would benefit from the advanced immunotherapy treatment. We further divided the radiomic metrics into three functional characteristics related to Size, Shape, Texture, and computed PCs in each category. The ensemble feature signatures in each of the categories were used in a logistic regression model to predict treatment response (1 year). We identified the benefit of radiomic feature extraction for the prediction of durable survival in R/R/DLBCL. We found that the Shape-based PCs showed comparable performance to MTV (see Table 2 ). Furthermore, they continue to complement MTV for patients with a smaller tumor burden (≤ median MTV) (see Supplementary Table S7 ). It has been shown that size-based features, including MTV-related metrics, are often prognostic for a patient’s outcome, as previously reported by us and others (12, 13, 28). A few other studies have associated the shape of lesions with aggressive phenotypes and was shown to have poor prognosis in oncological diseases (3, 29, 30). In our study, we observed that shape-based predictors select patients with large irregularly featured nodules, while more round features are associated with low features that are associated with better prognostic cases (see Figure 4 ).

The current clinical consensus assessment criteria (Cheson and Lugano classifications) (22, 31) do not always accurately capture the disease condition nor can they predict disease outcome, creating a need to develop better biomarkers. Characterizing subtle physiological changes observed on radiological imaging using quantitative metrics has shown enormous promise in developing models that can be related to patient outcomes (14, 16, 32). Our metrics may be capable of capturing nuances across lymphatic and extranodal lesions. Most extranodal lymphoma lesions are restricted to a single organ but may have infiltrative bone involvement, making them multifocal (31, 33). It is interesting that radiologists usually assess PET/CT images by characterizing the intensity and lesion shape characteristics to discern true cancerous lesions from benign lesions, an essential step in assessing the disease burden or MTV in lymphoma. Specifically, it has been reported that normal lymph nodes are often elongated in shape and have uniform non-thickened nodal cortices and a preserved fatty hilum (31). Similar observations have been reported for related shape-based characteristics and their ability to predict outcomes in oncological diseases (32, 34).

Lymphomatous lesions may manifest in a variety of appearances but are typically classified based on their cellular architecture (follicular or diffuse) and morphology (small or large) (35). Beyond tumor metabolic activity, the shape of the lesions observed on PET/CT also plays a diagnostic role in lymphoma (36). There have been many studies assessing the role of axi-cel therapy in DLBCL through immune dysregulation (10), but the role of advanced imaging analytics has not been adequately characterized.

Most quantitative assessments in lymphoma utilize MTV, which has been shown to be prognostic (12, 37) but is still not clinically adopted. This metric provides a gross measurement of disease burden in a patient and does not assess nor discriminate based on the subtle characteristics of the lesions. Our study independently assessed lesion changes (on the pixel level) in a 3D region in FDG-PET images that were matched to corresponding CT regions, revealing subtle details with higher levels of contrast compared to PET imaging alone (23). Recently, MTV combined with LDH was associated with progression-free survival and the radiomic metrics of these patients were associated with a complete response (19). In predictive performance, these radiomic features demonstrated comparable predictive performance within our cohort (see Table 2 , Figure 2 ).

Our finding provides a methodology for the use of radiological imaging-based quantitative metrics to assess patients’ response to immunotherapy (axi-cel) in DLBCL. Our systematic approach allows us to characterize the lesions’ shape irregularities and heterogeneity (including the microenvironment). It is well recognized that radiomic metrics are dependent on system-level variables such as scan parameters and image quality but despite these, the metrics can discriminate (38). Additionally, it is often difficult to explicitly relate a given radiomic feature to human observable characteristics, except for (mostly) size and shape metrics. Our study found that non-circular shaped lesions (high Shape PC) had a poor prognosis compared to close-to-circular-shaped lesions (low Shape PC) (see Figure 4 ).

It should be noted that shape-based metrics have been shown to have better reproducibility compared to texture metrics (39, 40). A recent study (20), found that a radiomic PET signature (using four features) achieved an AUC of 0.73 for predicting clinical benefit, outperforming metabolic tumor volume (AUC of 0.66).

Furthermore, our study needs secondary validation to provide prognostic benefits to patients treated with CAR-T for DLBCL. We retrospectively obtained diverse patient data for this study at our center and from a consortium that conducted the preliminary trial. Despite the efforts, the converged samples used in the study were limited, and we require additional independent samples to validate our findings. To mitigate and minimize study bias, we performed cross-validation to estimate predictor performance and reported our findings as an average (test) metric over multiple samplings. Our study provides several potential signatures (features) that need further validation to establish their use in the clinical setting and provide benefits for our patients. Our study follows the recommendations from the Standards for Reporting Diagnostic Accuracy Studies and Radiomics Quality Score to promote the reproducibility of clinical studies.

This study evaluates the ability of radiological imaging characteristics obtained from PET/CT images to assess DLBCL patients’ treatment outcomes. The quantified shape-based radiomic features (and some texture-based features) observed on PET images predicted patient’s response to axi-cel (CAR-T) immunotherapy treatment and were comparable to metabolic tumor volume in this regard.