802 Background: Predicting bladder cancer outcomes remains challenging due to pronounced tumor heterogeneity, a challenge that is inadequately addressed by the limited prognostic power of single-modality models. Although mitotic activity and tumor-infiltrating lymphocytes are both established prognostic pathology hallmarks in bladder cancer, their combined value for prognostic modeling has not been previously assessed. Methods: We present HyMiTiM (Hypergraph fusion of Multimodal Image features with Tumor Infiltration and Mitotic activity informed), a multimodal framework that integrates computed tomography (CT) and whole-slide image (WSI) features to predict bladder cancer prognosis. The first branch, HyMi, captures intratumoral heterogeneity by decomposing multiscale inputs into tumor-centered CT subvolumes and WSI patches. A lightweight hypergraph models cross-scale feature interactions between those fine-grained elements, and an attention pooling head aggregates them into a patient-level representation. To enhance biological interpretability, two additional WSI-based branches quantify pathomic phenotypes associated with immune infiltration lymphocytes (TIL) and mitotic activity, anchoring survival prediction in established bladder cancer morphological hallmarks. The final HyMiTiM risk score is derived by applying Platt scaling to the three branch outputs, producing a unified prognostic score. HyMiTiM was trained on the Qingdao Provincial Hospital cohort (QDPH, N = 436) and externally tested on the TCGA cohort (N = 71) and the Zhejiang Provincial Hospital (ZJPH, N = 43) cohort using matched CT scans and WSI. Results: In univariable analysis, HyMiTiM was prognostic of progression-free survival (PFS) on the TCGA cohort (HR = 2.26, 95% CI, 1.08–4.75, p = 0.0272) and ZJPH cohort (HR = 3.82, 95% CI, 1.44–10.17, p = 0.004). In terms of C-indices on the two external test cohorts, HyMiTiM (0.70 and 0.75) outperformed the three individual input branches: HyMi score (0.68 and 0.66), TIL score (0.67 and 0.63) and mitotic activity level (0.61 and 0.62). Conclusions: By unifying deep embeddings across radiology CT and pathology WSI as well as interpretable morphology phenotypes within WSI, HyMiTiM provides an interpretable multimodal fusion framework for bladder cancer prognosis.
Maurya et al. (Sun,) studied this question.