Magnetic resonance imaging (MRI) radiomics as predictor of clinical outcomes to neoadjuvant immunotherapy in patients with muscle invasive bladder cancer undergoing radical cystectomy.

1567 Background: Muscle-invasive bladder carcinoma (MIBC) is a deadly disease, for which we pioneered the use of neoadjuvant immune-checkpoint inhibitors (ICI) in a clinical trial (PURE-01, NCT02736266) testing 3 cycles of neoadjuvant pembrolizumab before radical cystectomy (RC). The objective of this study is to assess the ability of radiomic features extracted from a robust MRI processing pipeline to predict the pathological response to neoadjuvant pembrolizumab. Methods: A total of 120 patients (pts) with MIBC (102M/18 F), with median age of 68 years, a clinical stage T2N0 (n = 53; 44%) or T3-4N0 (n = 67; 56%), who were enrolled in PURE-01 study were analyzed. Patients had matched pre- and post-ICI MRIs, and tumors were segmented on both T2w images by GU radiologists. The MRI signal intensities were standardized by N4-bias field correction and robust z-scores. IBSI-compatible pyCERR software was used to extract radiomics features. A total of 289 radiomic features, including shape, first-order statistics, and higher-order textures, were analyzed for associations with pathological complete response (pCR at RC). An additional association was also investigated for major response groups, i.e., CR and partial response (PR, i.e. downstaging to ypT≤1N0) versus no response (NR). We employed Elastic Net, a machine learning technique that blends the strengths of Lasso and Ridge regression and is particularly effective for datasets with many correlated features such as in our study. The endpoint was modeled by training Elastic Net logistic regression models separately for pre- and post-ICI MRI features, as well as clinical T-stage. Models were evaluated on a 30% held-out test set using ROC curves (AUC). Results: For pCR, the best-performing model included four post-ICI MRI features: shape (flatness) and texture features from Gray Level Co-occurrence Matrix (GLCM: homogeneity, sum average, and sum entropy), and had a test AUC of 0.83 (95%CI: 0.66 - 0.99). Separate models fit on pre-ICI MRI features selected two important pre-ICI MRI features: shape (surface-to-volume ratio) and first order (robust mean absolute deviation), but the overall performance was lower than post-ICI models (test AUC 0.66; 95%CI: 0.42 - 0.89). For major response assessment, the best-performing model included two post-ICI MRI features: shape (flatness) and texture (GLCM sum average) and had a test AUC of 0.92 (95%CI: 0.8-1.0). Conclusions: This is one of the first machine learning models using MRI radiomics to predict neoadjuvant immunotherapy response in pts with MIBC. These results could be instrumental for improving the way we can predict the pathological response in these pts. Clinical trial information: NCT02736266 .