Radiomics signature for dynamic monitoring of tumor inflamed microenvironment and immunotherapy response prediction

Abstract

Background: The efficacy of immune checkpoint inhibitors (ICIs) depends on the tumor immune microenvironment (TIME), with a preference for a T cell-inflamed TIME. However, challenges in tissue-based assessments via biopsies have triggered the exploration of non-invasive alternatives, such as radiomics, to comprehensively evaluate TIME across diverse cancers. To address these challenges, we develop an ICI response signature by integrating radiomics with T cell-inflamed gene-expression profiles. Methods: We conducted a pan-cancer investigation into the utility of radiomics for TIME assessment, including 1360 tumors from 428 patients. Leveraging contrast-enhanced CT images, we characterized TIME through RNA gene expression analysis, using the T cell-inflamed gene expression signature. Subsequently, a pan-cancer CT-radiomic signature predicting inflamed TIME (CT-TIME) was developed and externally validated. Machine learning was employed to select robust radiomic features and predict inflamed TIME. The study also integrated independent cohorts with longitudinal CT images, baseline biopsies, and comprehensive immunohistochemistry panel evaluation to assess the pan-cancer biological associations, spatiotemporal landscape and clinical utility of the CT-TIME. Results: The CT-TIME signature, comprising four radiomic features linked to a T-cell inflamed microenvironment, demonstrated robust performance with AUCs (95% CI) of 0.85 (0.73 to 0.96) (training) and 0.78 (0.65 to 0.92) (external validation). CT-TIME scores exhibited positive correlations with CD3, CD8, and CD163 expression. Intrapatient analysis revealed considerable heterogeneity in TIME between tumors, which could not be assessed using biopsies. Evaluation of aggregated per-patient CT-TIME scores highlighted its promising clinical utility for dynamically assessing the immune microenvironment and predicting immunotherapy response across diverse scenarios in advanced cancer. Despite demonstrating progression disease at the first follow-up, patients within the inflamed status group, identified by CT-TIME, exhibited significantly prolonged progression-free survival (PFS), with some surpassing 5 months, suggesting a potential phenomenon of pseudoprogression. Cox models using aggregated CT-TIME scores from baseline images revealed a statistically significant reduction in the risk of PFS in the pan-cancer cohort (HR 0.62, 95% CI 0.44 to 0.88, p=0.007), and Kaplan-Meier analysis further confirmed substantial differences in PFS between patients with inflamed and uninflamed status (log-rank test p=0.009). Conclusions: The signature holds promise for impacting clinical decision-making, pan-cancer patient stratification, and treatment outcomes in immune checkpoint therapies.