The AI-MIRACLE study investigated whether artificial intelligence-based analysis of multiparametric MRI could predict pathological response to neoadjuvant pembrolizumab in patients with MIBC. The study included 112 patients from the PURE-01 trial who underwent MRI before and after immune-checkpoint inhibitor therapy, followed by radical cystectomy. The imaging analysis was performed using a multi-institutional workflow, with MRI acquisition in Italy and centralized image analysis in the United States. The models incorporated different types of MRI-derived information, including radiomic features extracted from T2-weighted images, dynamic contrast-enhanced MRI parameters, diffusion-weighted imaging features, and deep learning-derived features.
The most relevant finding was that post-treatment MRI features predicted pathological major response with high accuracy. Pathological major response was defined as residual disease lower than ypT2N0, a clinically meaningful endpoint because it may identify patients with substantial tumor downstaging after immunotherapy. The best-performing models achieved an AUC of 0.96. This was obtained both with a model combining radiomics and dynamic contrast-enhanced MRI features and with a simpler model based on radiomics alone.
For pathological complete response, defined as ypT0N0, the best model was also based on post-treatment radiomics and achieved an AUC of 0.86. Although prediction of complete response remains more challenging than prediction of major response, this result is clinically important because complete response is the key endpoint for future bladder-preservation strategies. Notably, models based on imaging features outperformed benchmark models using clinical variables alone, including clinical stage, tumor mutational burden, and PD-L1 expression.
The study also showed that the most informative MRI features were not limited to simple tumor size. Tumor shape, texture, and vascular-related parameters contributed to response prediction, suggesting that post-treatment MRI may capture biological changes induced by immunotherapy that are not evident from conventional clinical assessment alone.
Several limitations should be acknowledged. This was a retrospective analysis within a clinical trial cohort, and the models require external validation. In addition, the workflow relied on expert image segmentation and centralized analysis, which may not yet reflect routine practice.
However, the results of AI-MIRACLE provide an important proof of concept: automated analysis of post-immunotherapy MRI may support response-adapted management in MIBC. In the future, such tools could be integrated with cystoscopy, TURBT, cytology, and molecular biomarkers to improve the selection of patients for bladder-sparing trials.
Written by: Giorgio Brembilla,1,2 Andrea Necchi2,3
- Department of Radiology, IRCCS San Raffaele Hospital, Comprehensive Cancer Center, Milan, Italy;
- Vita-Salute San Raffaele University, Milan, Italy;
- Department of Medical Oncology, IRCCS San Raffaele Hospital, Comprehensive Cancer Center, Milan, Italy.