AI-MIRACLE: Artificial Intelligence and MultIpaRAmetric MRI Predict CLinical OutcomEs to Neoadjuvant Immunotherapy in Patients with Muscle-invasive Bladder Cancer Undergoing Radical Cystectomy.
Authors
Affiliations (8)
Affiliations (8)
- Department of Medical Oncology, IRCCS San Raffaele Hospital, Milan, Italy; Department of Medical Oncology, Vita-Salute San Raffaele University, Milan, Italy. Electronic address: [email protected].
- Department of Medical Oncology, Vita-Salute San Raffaele University, Milan, Italy; Department of Radiology, IRCCS San Raffaele Hospital, Milan, Italy.
- Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center, New York, USA.
- Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, USA.
- Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA.
- Department of Radiology, IRCCS San Raffaele Hospital, Milan, Italy.
- Department of Medical Oncology, IRCCS San Raffaele Hospital, Milan, Italy.
- Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, USA; Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA. Electronic address: [email protected].
Abstract
Neoadjuvant immune-checkpoint inhibitors (ICIs) in muscle-invasive bladder cancer (MIBC) were tested in patient's ineligible for cisplatin-based chemotherapy. The PURE-01 trial (NCT02736266) evaluated three courses of pembrolizumab before radical cystectomy (RC). We developed AI-MIRACLE, an international study assessing artificial intelligence (AI) and multiparametric magnetic resonance imaging (mpMRI) for predicting treatment response. This multi-institutional study included data acquisition in Italy, and centralized analysis in the United States. Among 112 PURE-01 patients, pre- and post-ICI MRIs were analyzed. T2-weighted signal intensities were standardized for radiomics (Image Biomarker Standardization Initiative-compatible Python-based Computational Environment for Radiological Research (pyCERR)) and deep feature extraction (AI-BLADE toolbox using VGG19). Diffusion-weighted (DW) and dynamic contrast-enhanced (DCE) MRI data underwent model-based analysis. Supervised machine learning algorithms (elastic net, random forest) were trained and cross-validated to predict pathological major response (pMR:ypT<2N0 residual disease) and pathological complete response (pCR: ypT0) pathological response. The predictive models using post-ICI mpMRI with either a combination of radiomics and DCE-derived features or radiomics alone achieved the same high accuracy, with an area under the receiver operating characteristic curve (AUC) of 0.96 for pMR. A shape-based radiomic model achieved an AUC of 0.86 for predicting pCR. These models outperformed benchmark models based on clinical predictors. Shape-based radiomics, DCE-derived features, and deep features may serve as noninvasive imaging biomarkers for predicting response to neoadjuvant pembrolizumab in MIBC. This imaging-based approach provides a non-invasive assessment of treatment response following neoadjuvant immunotherapy, which may help inform bladder-preserving management decisions prior to definitive surgery.