Non-invasive characterization of intratumoral CD8+ T cells using standard-of-care (SOC) CT and ⁸⁹Zr-crefmirlimab berdoxam PET (CD8-PET) radiomic signature in solid tumors.

Abstract

2614 Background: Noninvasive estimation of intratumoral CD8+ T-cell density using radiomics may enhance comprehensive immune profiling and support immuno-oncology development and treatment decisions. We evaluated the feasibility of using a multimodality radiomic approach incorporating SOC CT and CD8-PET/CT to characterize CD8+ T-cell density in solid tumor lesions. Methods: 71 soft-tissue lesions from 52 patients with solid tumors enrolled in a Phase II iCorrelate trial (NCT03802123) were retrospectively analyzed. Patients received a pre-treatment CD8-PET/CT scan (ImaginAb, Inc) before initiating SOC immunotherapy (IOT) and a second CD8-PET/CT scan on-treatment 4-8 weeks later, with selection of biopsied tumor lesions of known CD8+ T-cell density for radiomic analysis. Radiomic features were extracted from volumetric segmentations of these tumor lesions using 3D Slicer on Diagnostic contrast CT (DCT, n=27), CT attenuation correction (CTAC; n=71), and CD8 PET (n=71) images. Features were integrated with acquisition parameters, lesion location, and adenopathy status. Feature selection used maximum relevance–minimum redundancy and variance inflation factor methods. Elastic-Net classifiers were trained to predict binarized (median split at the 328 cells/mm 2 level) CD8+ T-cell density, with hyperparameter optimization via grid search and 5 fold Cross-Validation (CV). In each iteration of the CV, 1 fold (20%) was left out and models were trained on the remaining 4 folds (80%). Model performance was assessed using AUC and F1 score. Results: In the test set, Multimodality models (CTAC+DCT+PET) outperformed single-modality (AUC up to 0.94 vs. 0.80), with improved AUC and F1 scores and robust performance across endpoints. Single-modality models achieved AUCs of 0.76 (F1 0.55) for DCT, 0.80 (F1 0.72) for CTAC, and 0.80 (F1 0.70) for PET. The combined CTAC+DCT+PET model demonstrated superior performance with a test AUC of 0.85 (F1 0.79) and training AUC of 0.94 (F1 0.89). Application of previously reported DCT radiomic weights (Sun et al.) to this cohort yielded a lower AUC of 0.68, compared with the multimodality model. Conclusions: This study demonstrates robust radiomics-based characterization of intratumoral CD8+ T-cell density in patients with solid tumors is feasible using non-invasive multi-modality approach utilizing CD8-PET and SOC CT scans. These findings may lead to further clinical adoption and non-invasive monitoring of the tumor immune microenvironments with CD8-PET and SOC CT scans using radiomics, informing IOT decision-making, and enhancing patient stratification in both clinical development and routine oncology practice. Validation in larger, multicenter cohorts is still required to confirm generalizability and robustness of these radiomic signatures. Clinical trial information: NCT03802123 .