MRI radiomics to predict outcome of neoadjuvant chemotherapy in patients with muscle invasive bladder cancer undergoing radical cystectomy.

Abstract

4581 Background: Cisplatin-based neoadjuvant chemotherapy (NAC) before radical cystectomy (RC) is the standard of care in patients with muscle-invasive bladder cancer (MIBC). Although the administration of NAC for MIBC has increased over the years, it still does not meet actual patient's needs, particularly in cT2 BC, for which it is currently recommended in clinical guidelines. Indeed, with the development of new cytotoxic and targeted therapies, large ongoing prospective studies have been designed to test their efficacy either alone or in combination in the neoadjuvant setting. Multidisciplinary management is critical in this disease setting; including advanced imaging to assess response to treatment and outcome correlations. Despite the promising applications of radiomics in MIBC treatment outcome assessment, challenges remain, including successful harmonizing of imaging data, which impacts the consistency of radiomic features. The objective of the study is to assess the ability of radiomic features extracted from a robust magnetic resonance imaging (MRI) processing pipeline to predict the outcome of NAC prior to RC in patients with MIBC. Methods: A total of 105 MIBC patients (67M/38F), median age (65), clinical stage 2 (77), 3(28)) who were treated with NAC (cisplatin-based therapy) and underwent RC were included in this study. All patients underwent preNAC MRIs using the standard acquisition protocol. Tumors were segmented on T2w, T1w, and post contrast-T1w images by GU radiologists. To standardize MRI intensity values across scans, preprocessing steps were required to ensure comparability between patients. After N4-bias field correction of image intensities, images were standardized to robust z-scores using median and mean absolute deviation of intensities within respective regions of interest. 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 the overall survival (OS, time between RC to death) as an outcome. To identify features associated with OS, we trained an Elastic Net Cox regression model for each MRI sequence, with performance evaluated by concordance index (c-index) on a 30% held-out test set. Results: The same shape feature (major axis length) from post contrast-T1w and T2w images was selected as important by the elastic net with test set c-index of 0.55 [0.42 – 0.67) and 0.56 [0.42 – 0.70], respectively. Kaplan-Meier method further confirmed the significance of this feature (p < 0.05) for OS risk stratification, using the median feature value as the cutoff point. Conclusions: The study demonstrated the value of radiomics in predicting survival to NAC with MIBC which can be further validated with a larger independent cohort. MRI radiomics may be an additional tool for prognostication.