Predicting survival outcomes in CAR-t treated B-cell lymphoma by combining PET/CT imagingand clinical features Free

Background Chimeric antigen receptor T-cell (CAR-T) therapy has emerged as a transformative option for relapsed/refractory (R/R) B-cell lymphomas. However, early identification of patients at higher risk of progression or death remains essential to optimize post-CAR-T management. Here, we aimed to assess whether quantitative imaging features from [¹⁸F]FDG PET/CT, along with clinical and molecular variables, could improve prediction of overall survival (OS) and progression-free survival (PFS) at pre-lymphodepletion (PreLD) and day +30 post-infusion.

Methods We retrospectively analyzed 109 consecutive patients diagnosed with diffuse large B-cell lymphoma (DLBCL,n = 88), primary mediastinal B-cell lymphoma (PMBCL, n = 13) or high-grade B-cell lymphoma (HGBCL, n = 8), treated with anti-CD19 CAR-T cells between 2019–2024 at two centres. [¹⁸F]FDG PET/CT scans were acquired at PreLD and +30 days (n = 60). All metabolically active lesions were manually segmented by an imaging technician with over 5 years of experience and supervised by a nuclear medicine physician with more than 20 years of experience. Radiomics features and conventional PET metrics were extracted. Imaging features were aggregated at patient level, and delta features were calculated when both timepoints were available. In additional binary variables was created to flag patients with complete metabolic response (CMR, n = 9) and imaging variables were imputed with cero values. Clinical and molecular data, including CAR-T expansion metrics and relevant genetic data, were included.

Survival models for OS and PFS were developed using Cox proportional hazards regression with leave-one-out cross-validation. At each timepoint, three model types were assessed: clinical/molecular only, imaging only, and combined. Feature selection was based on Cox univariate analysis; clinically relevant variables of proven prognostic value (CRS, ICANS, bulky disease_PreLD, ferritin_PreLDratio, LDH_PreLD ratio, C-reactive protein_PreLDratio) were systematically retained and excluded from selection.

Results Median age was 62 (range, 21–79) years; 39.5 % were female. At Pre-LD, most patient had stage IV disease (n = 63, 57.8%), 49 (44.95%) had an International Prognostic Index (IPI) scores of 3–5, and 27 (24.8%) had bulky disease. The median number of prior treatment lines was 2 (range, 1–10) and 79.8% received bridging therapy. CAR-T product received included axicabtagene ciloleucel (n = 77, 70.6%), tisagenlecleucel (n = 25, 22.9%), and CYTB323A12101 (n = 7, 6.4%) in a clinical trial. Median OS was 8.96 (range 0.20–54.67) months, and median PFS was 3.05 (range 0.90–53.87) months.

Imaging-only, clinical-only and combined models for OS prediction showed modest performance at PreLD (C-index: 65.4%, 68.9%, 68.7%, respectively) and improved at day +30 (79.5%, 76.6%, 79.9%, respectively). Key PreLD imaging predictors included SUV_3Dpeak, and a radiomic feature reflecting irregular tumor morphology. Relevant clinical variables were ECOG_PreLD, IPI_PreLD, and patient age. At day +30, the strongest imaging predictors were SUV_std, SUV_p25, and a radiomic tumor heterogeneity-related feature, along with ECOG_PreLD, IPI_PreLD, and ferritin_PreLD ratio.

PFS models followed a similar trend, with imaging only, clinical only and combined models showing modest performance at PreLD (C-index: 67.7%, 66.3%, 70.2% respectively) and marked improvement at day +30 (78.7%, 67.8%, 77.9%, respectively). PreLD imaging predictors included SUV_3Dpeak, radiomic features reflecting irregular tumor morphology and the minimum intensity value, as well as ECOG_PreLD, IPI_PreLD, C-reactive protein_PreLDratio, and the CAR-T type. At day +30, imaging features such as SUV_std, SUV_p25 and tumor morphology were the most informative.

Delta metrics were evaluated for OS and PFS models; however, they did not show significant prognostic value. Instead, the persistence and degree of residual metabolic activity, along with tumor heterogeneity, were more strongly associated with survival outcomes.

ConclusionsQuantitative PET imaging features retain prognostic value for survival in CAR-T treated R/R B-cell lymphoma patients. Although PreLD models demonstrated limited prognostic performance, a substantial improvement was observed at day +30, with imaging-based models outperforming clinical ones. These findings support the integration of PET-based radiomics for early risk stratification to guide early individualized follow-up strategies.