Magnetic resonance imaging radiomic phenotypes stratify response and define immune states in very high‐risk, nonmuscle–invasive bladder cancer treated with tislelizumab plus nanoparticle albumin‐bound paclitaxel

BACKGROUND: Patients with very high-risk (VHR), nonmuscle-invasive bladder cancer (NMIBC) who are ineligible for or decline radical cystectomy face limited treatment options. Immune checkpoint inhibitor-based regimens have demonstrated promise but lack reliable biomarkers to guide patient selection. The objective of this study was to identify magnetic resonance imaging-derived radiomic phenotypes predictive of treatment response to immune checkpoint inhibitor-based therapy. METHODS: The authors retrospectively analyzed two prospective cohorts of patients with extensive VHR NMIBC: a single-center derivation cohort from the TRUCE-02 trial (n = 59) and a multicenter validation cohort from the TRUCE-22 trial (n = 37). All patients received tislelizumab and nanoparticle albumin-bound paclitaxel every 3 weeks. Complete response (CR) was assessed at 3 months. Radiomic features were extracted from baseline, T2-weighted images, and a hybrid framework integrating supervised feature selection with unsupervised clustering was used to identify imaging phenotypes. Associations with CR, progression-free and overall survival, and immune-related transcriptional signatures were evaluated. RESULTS: Three radiomic clusters were identified: immune-excluded (cluster 0), immune-intermediate (cluster 1), and immune-inflamed (cluster 2). In the derivation cohort, cluster 2 had a significantly higher CR rate (82.8%) than cluster 0 (21.4%; p = .001), a pattern that was validated in TRUCE-22 (83.3% vs. 22.2%; p = .008). The cluster-only model achieved the highest predictive performance for CR in both the derivation (area under the receiver operating characteristic curve, 0.77) and validation (area under the receiver operating characteristic curve, 0.78) cohorts. Cluster 2 had better progression-free and overall survival. Transcriptomic and immunofluorescence analyses revealed functional immune suppression in cluster 0, including reduced CD8-positive, phosphorylated STAT1-positive T cells. CONCLUSIONS: In patients with very high-risk NMIBC, magnetic resonance imaging-derived radiomic phenotypes can be used to stratify responses to immunochemotherapy and capture functional immune states. Noninvasive radiomic profiling may guide precision treatment selection in patients who are unfit for radical cystectomy, and prospective validation is warranted before clinical adoption.