Radiomic models based on magnetic resonance imaging predict the spatial distribution of CD8+ tumor-infiltrating lymphocytes in breast cancer

Abstract

Infiltration of CD8+ T cells and their spatial contexture, represented by immunophenotype, predict the prognosis and therapeutic response in breast cancer. However, a non-surgical method using radiomics to evaluate breast cancer immunophenotype has not been explored. Here, we assessed the CD8+ T cell-based immunophenotype in patients with breast cancer undergoing upfront surgery (n = 182). We extracted radiomic features from the four phases of dynamic contrast-enhanced magnetic resonance imaging, and randomly divided the patients into training (n = 137) and validation (n = 45) cohorts. For predicting the immunophenotypes, radiomic models (RMs) that combined the four phases demonstrated superior performance to those derived from a single phase. For discriminating the inflamed tumor from the non-inflamed tumor, the feature-based combination model from the whole tumor (RM-wholeFC) showed high performance in both training (area under the receiver operating characteristic curve [AUC] = 0.973) and validation cohorts (AUC = 0.985). Similarly, the feature-based combination model from the peripheral tumor (RM-periFC) discriminated between immune-desert and excluded tumors with high performance in both training (AUC = 0.993) and validation cohorts (AUC = 0.984). Both RM-wholeFC and RM-periFC demonstrated good to excellent performance for every molecular subtype. Furthermore, in patients who underwent neoadjuvant chemotherapy (n = 64), pre-treatment images showed that tumors exhibiting complete response to neoadjuvant chemotherapy had significantly higher scores from RM-wholeFC and lower scores from RM-periFC. Our RMs predicted the immunophenotype of breast cancer based on the spatial distribution of CD8+ T cells with high accuracy. This approach can be used to stratify patients non-invasively based on the status of the tumor-immune microenvironment.