Introduction: We previouslyreportedthe feasibility of predicting treatment outcome of chimeric antigen receptor modified T-cell (CAR-T) therapy for patients (pts) with relapsed or refractory large B-cell lymphomas (r/r LBCL) from pretreatment diagnostic imaging studies. This method used deep-learning (DL)-based image analysis for lesion-level response prediction and estimated patient-level outcomes from lesion-level predictions by rule-based reasoning (Tong Y, et al. PLoS ONE 2023;18(7):e0282573). To further test the prognostic validity of this method, we analyzed baseline (pre-CAR-T infusion) FDG-PET (PET) and low-dose CT (LD-CT) images performed as part of the JULIET trial with investigators blinded to pt treatment outcomes. The JULIET trial is a phase 2 study of tisagenlecleucel, a CD19-directed CAR-T therapy, in adult pts with r/r LBCL (NCT02445248). Here, we compare this approach with serum LDH and secondary International Prognostic Index (sIPI), which are generally accepted prognostic markers for LBCL treatment outcome.

Methods: Pre-infusion imagesfrom 102 adult pts with r/r LBCL who were treated with tisagenlecleucel were evaluated. Image sets came from 27 hospitals in 10 countries and were acquired on 15 different model scanners from 3 leading manufacturers of diagnostic imaging equipment; 36 (35%) pt image sets were excluded from DL-based image analysis: 31 (30%) due to image low quality (26 LD-CT; 4 LD-CT+PET; 1 PET), 1 without nodal lesion, and 4 without metabolic confirmation of Month-12 response. Data from 3 contiguous whole-image slices through the mid-portions of nodal lesions on both PET and LD-CT images from each of the 66 pt image sets were analyzed using the previously described DL lesion-level model, without retraining, to predict treatment outcome. After analyzing image sets to generate predictions of outcome for each of the 66 evaluable pts, actual Month-12 pt outcomes were unblinded and grouped using protocol-specified radiologic response criteria as a Responder (met protocol-defined radiologic complete response [CR]) or a Non-responder (met protocol-defined radiologic partial response, stable disease or progressive disease [< CR]) at 12 months post CAR-T infusion. Actual Month-12 pt outcomes post-treatment (verified Responders, n = 13 [i.e., 20% CR]; verified Non-responders, n = 53 [i.e., 80% < CR]) were then compared with predictions obtained in blinded fashion from DL-based image analyses using 70% rule-based reasoning (i.e., if > 70% of lesions were predicted to respond by DL image analysis, the pt was predicted to be a Responder at Month-12; if < 70% lesions were predicted to respond, the pt was predicted to be a Non-responder at Month-12).

Results: For 66 evaluable pts, DL-based prediction of Responder (CR) status as the outcome at Month-12 had a sensitivity of 77% (correctly identified Responders) with specificity 51% and balanced accuracy 64% (balanced accuracy reported because of imbalance between number of Responders and non-Responders).

For comparison, serum LDH and sIPI score at enrollment were also evaluated as prognostic indices for 64 pts (2/66 pts excluded for missing LDH). Using serum LDH < 2 x upper limit of normal (ULN) to predict the outcome at Month-12 as Responder, sensitivity was 100% but specificity only 12% due to a high false positive rate; using LDH > 2 x ULN to predict outcome as Non-responder, sensitivity was only 12% (high false negative rate) with specificity 100%. Using sIPI > 2 to predict outcome as Non-responder at 12 months, sensitivity was 65% and specificity 62%.

Conclusions: Prediction of CAR-T treatment outcome from pretreatment images using DL-based image analysis for lesion-level response prediction and rule-based reasoning for patient-level response estimation is feasible. Considering the challenges stemming from data heterogeneity and the small number of pts in this study, this approach showed a generalizable performance with the accuracy of patient-level predictions similar to earlier results obtained from our single center study. With continued refinement and addition of clinical covariates, this approach has the potential to provide clinically useful information in advance of CAR-T therapy.

This content is only available as a PDF.
2025
Sign in via your Institution