--- title: "CHIMERA Challenge Task 2 and 3 response subtypes classification and progression prediction in bladder cancer Netherlands; First standardized multimodal benchmark with 368 patients" sdDatePublished: "2026-09-10T04:10:00Z" source: "https://arxiv.org/pdf/2609.09510" topics: - name: "health care provider" identifier: "medtop:20000278" - name: "medical research" identifier: "medtop:20000737" - name: "scientific innovation" identifier: "medtop:20000736" - name: "artificial intelligence" identifier: "medtop:20001298" - name: "cancer" identifier: "medtop:20000447" locations: - "Tokyo" - "Toronto" - "Stavanger" - "Fort Worth" - "San Diego" - "Tampere" - "Bydgoszcz" - "Udine" - "Yerevan" - "Hefei" - "Suzhou" - "Beiyuan" - "Coventry" - "Gelderland" - "North Brabant" - "Darmstadt" - "Poland" - "Italy" - "United States" - "Portugal" - "Canada" - "United Kingdom" - "Armenia" - "China" - "Germany" - "Japan" - "Finland" - "Netherlands" - "Norway" --- CHIMERA Challenge Task 2 and 3 response subtypes classification and progression prediction in bladder cancer Netherlands; First standardized multimodal benchmark with 368 patients CHIMERA Challenge Task 2 and 3: Response Subtypes Classification and Progression Survival Prediction in Bladder Cancer Patients using Multimodal Datasets Graphical Abstract CHIMERA Challenge Task 2 and 3: Response Subtypes Classification and Progression Survival Prediction in Bladder Cancer Patients using Multimodal Datasets Catherine Chia, Tongjie Wang, Robert Spaans, Maryam Mohammadlou, Far- bod Khoraminia, J. Alberto Nakauma-González, Adam Kowalewski, Parandzem Khachatryan, Domingos Oliveira, Khrystyna Faryna, CHIMERA Challenge Con- sortium, Marlies Wakkee, Sita Vermeulen, Tahlita Zuiverloon, Nadieh Khalili arXiv:2609.09510v1 [eess.IV] 8 Sep 2026 Highlights CHIMERA Challenge Task 2 and 3: Response Subtypes Classification and Progression Survival Prediction in Bladder Cancer Patients using Multimodal Datasets Catherine Chia, Tongjie Wang, Robert Spaans, Maryam Mohammadlou, Far- bod Khoraminia, J. Alberto Nakauma-González, Adam Kowalewski, Parandzem Khachatryan, Domingos Oliveira, Khrystyna Faryna, CHIMERA Challenge Con- sortium, Marlies Wakkee, Sita Vermeulen, Tahlita Zuiverloon, Nadieh Khalili • First standardized multimodal benchmark for bladder cancer. • Cohort shift and missing data limit multimodal model general- ization. • Cross-model errors reveal consensus-hard patients associated with T1 substage. CHIMERA Challenge Task 2 and 3: Response Subtypes Classification and Progression Survival Prediction in Bladder Cancer Patients using Multimodal Datasets Catherine Chiaa,c,d,f,1,∗, Tongjie Wanga,b,1,∗∗, Robert Spaansa,e,f, Maryam Mohammadloug, Farbod Khoraminiab, J. Alberto Nakauma-Gonzálezb, Adam Kowalewskia,h,i, Parandzem Khachatryanj, Domingos Oliveirak, Khrystyna Farynaa,f, CHIMERA Challenge Consortiumm, Marlies Wakkeec, Sita Vermeulenl, Tahlita Zuiverloonb, Nadieh Khalilia,f aDepartment of Pathology, Research Institute for Medical Innovation, Radboud University Medical Center, the Netherlands bDepartment of Urology, Erasmus MC Cancer Institute, Erasmus University Medical Center, the Netherlands cDepartment of Dermatology, Erasmus MC Cancer Institute, Erasmus University Medical Center, the Netherlands dDepartment of Pathology, Erasmus University Medical Center, the Netherlands eDepartment of Urology, Canisius Wilhelmina Hospital, the Netherlands fOncode Institute, the Netherlands gCenter for Cancer Eradication Research, Faculty of Medicine and Health Technology, Tampere University and TAYS Cancer Center, Tampere University Hospital, Finland hFaculty of Medicine, Bydgoszcz University of Science and Technology, Bydgoszcz, Poland iDepartment of Tumor Pathology, Oncology Centre Prof. Franciszek Łukaszczyk Memorial Hospital, Bydgoszcz, Poland jYerevan State Medical University, Armenia kResearch and Development Unit, IMP Diagnostics, Portugal lGenetic epidemiology, Radboud University Medical Center, The Netherlands m17 Affiliations. See Section CHIMERA Challenge Consortium Abstract High-risk non–muscle-invasive bladder cancer (HR-NMIBC) carries a substan- tial risk of recurrence and disease progression with poor prognosis. Current clinical risk stratification is inadequate at predicting these prognoses. A com- bination of clinicopathological parameters and molecular subtyping showed po- tential to improve risk stratification. We hypothesized that integration of mul- timodal information from clinicopathological, histopathological and molecu- lar data has the potential to further improve patient risk stratification and prognostic prediction. While advancement in multimodal artificial intelligence (AI) is promising, progress remains constrained without standardized bench- mark datasets and systematic benchmarking. CHIMERA, a multimodal AI challenge was organized with two tasks targeting HR-NMIBC: Task BRS (Bacillus Calmette-Guérin (BCG) Response Subtypes) predicts molecular sub- types using histopathology and clinicopathological data, and Task Progres- sion models time-to-progression using histopathology, clinicopathological and RNA-sequencing data. A multimodal dataset comprising 368 patients was split into public training and hidden validation and test sets. A total of 159 submis- sions were made to Grand Challenge, with 13 top-performing models selected for benchmarking. These models achieved a weighted F1 of 0.73 (Task BRS) and a C-index of 0.68 (Task Progression). Benchmarking revealed task-dependent modality contribution. Histopathology compensated for missing pathology- derived variables in clinicopathological data in Task BRS, but Task Progres- sion models showed dependence on such variables. Cohort-stratified test per- formance degradation and sensitivity to missing data highlighted barriers to transportability, motivating missingness-aware architectures and validation in fully independent multi-institutional cohorts. CHIMERA represents the first benchmark for evaluating multimodal AI methods in bladder cancer and estab- lishes a foundation for future methodological advances in oncologic prediction modeling. Keywords: Multimodal AI, Artificial Intelligence, Bladder Cancer, Grand Challenge 1. Introduction High-risk non-muscle-invasive bladder cancer (HR-NMIBC) is managed with transurethral resection of the bladder tumor (TURBT) followed by adjuvant in- travesical bacillus Calmette–Guérin (BCG) instillations [1]. Although BCG can induce durable disease control in a subset of patients, ˜50% of patients experience recurrence and/or progression [2], resulting in intensified surveillance, repeated procedures, and difficult treatment-escalation decisions. In particular, delaying definitive treatment in patients unlikely to benefit from BCG can expose them to unnecessary toxicity and postpone potentially curative interventions. Tra- ditional risk stratification approaches rely on clinicopathological variables and population-level risk groups, yet individual outcomes are heterogeneous [3]. Recent molecular profiling studies have highlighted that poor outcomes after BCG are not random but are more common in biologically distinct molecular subtypes[4]. De Jong et al. identified and validated three clinically relevant BCG Response Subtypes (BRS 1–3) using transcriptomic profiling of HR-NMIBC. BRS3 tumors were associated with inferior recurrence-free and progression-free survival and with BCG failure [4]. Importantly, BRS stratification outperformed guideline-recommended clinicopathological risk stratification in an independent validation cohort, supporting the clinical relevance of these subtypes for treat- ment selection and trial design [4] Generating such profiles requires dedicated tissue-processing and sequencing workflows, together with validated computational pipelines and bioinformatics ∗Corresponding author. (email: catherine.chia@radboudumc.nl) ∗∗Corresponding author. (email: t.j.wang@erasmusmc.nl) 1These authors contributed equally to this work expertise; this process is time- and cost-intensive and requires allocation of finite tumor tissue for RNA extraction[5]. Consequently, RNA-sequencing (RNA-seq) based transcriptomic subtyping is not yet broadly embedded in routine pathol- ogy workflows, creating a translational gap between the potential prognostic value of molecular subtypes and their practical availability at the point of care. In contrast, hematoxylin and eosin (H&E) stained histopathology is routinely generated during diagnostic assessment and captures morphological phenotypes associated with underlying tumor biology. Computational pathology studies have demonstrated that molecular phenotypes, including transcriptomic pro- files and molecular subtypes, can be inferred from H&E whole slide images (WSIs) [6, 7]. This motivates the first bladder-cancer task in CHIMERA: to determine whether BCG response subtypes can be predicted from routine H&E WSIs together with structured clinicopathological information, thereby reducing reliance on RNA sequencing for subtype assignment. Even when transcriptomic measurements are available, molecular subtypes derived from bulk RNA-seq provide an incomplete representation of intratumor heterogeneity. Bulk expression profiles aggregate signals across heterogeneous mixtures of malignant, stromal, and immune-cell populations, potentially ob- scuring cell-state heterogeneity [8], and do not preserve the spatial organization of these components within the tissue [9]. Intratumor subtype heterogeneity is therefore relevant both to discrete subtype assignment and to clinical out- come: molecularly distinct tumor-cell states can coexist within a single blad- der tumor, and higher representation of the high-risk class 2a state was asso- ciated with worse progression-free survival in NMIBC [10]. Conversely, H&E histopathology preserves spatial tissue architecture and cellular morphology but does not directly measure gene-expression programs. RNA-seq and histopathol- ogy may therefore provide complementary views of disease: transcriptomics captures molecular activity, whereas histology captures spatial architecture and phenotypic organization. Structured clinical and treatment variables provide an additional layer of patient-level context. Prior multimodal studies integrat- ing histopathology with molecular data, including RNA-seq-derived features, have demonstrated the potential of such complementary information for cancer prognosis and risk stratification [11, 12]. Integrating these modalities may con- sequently improve individualized progression-risk estimation beyond any single data modality alone. Despite rapid progress in computational pathology and multimodal learning, robust assessment of such approaches remains difficult. Methodological com- parisons are frequently confounded by heterogeneous cohorts, endpoints, data modalities, and evaluation protocols [13, 14]. Multimodal models additionally face practical challenges related to cross-modal harmonization, incomplete in- puts, cohort-specific acquisition effects, and generalization across institutions and workflows. Moreover, algorithms developed and evaluated in isolated set- tings may exhibit bias or limited transportability when applied to independent populations [15, 16]. Community challenges provide an effective mechanism to crowd-source methodological solutions under a common evaluation framework, but clinically credible benchmarking requires prespecified evaluation procedures, 3 Training Validation Test Challenge Phases Benchmarking Public dataset Hidden Hidden Full data access Algorithm performance Leaderboards Final model submissions Analyses Participants Top teams Summary of algorithms Structured data:  missing values Structured data: missing variables Reduced models: removing modality- overlapping variables Error analysis Robustness CHIMERA Combining Histology, medIcal iMaging (radiology) and molEcular data for medical pRognosis and diAgnosis Figure 1: Overview of the CHIMERA Challenge phases and benchmarking workflow for blad- der cancer multimodal modeling (Task BRS and Task Progression). Participants developed models using the public training data and evaluated them on a hidden validation set. Final models were then tested on the hidden test set. Top-performing solutions were included in a six-part benchmarking analysis. strict separation between development and final testing, and transparent base- lines [14, 17]. To address these challenges, Combining Histology, medIcal iMaging (radi- ology) and molEcular data for medical pRognosis and diAgnosis (CHIMERA) was established as an international multimodal benchmarking challenge. Within HR-NMIBC, CHIMERA comprised two complementary tasks. Task 2, hereafter Task BRS, evaluated prediction of RNA-seq-based BCG response subtype from H&E WSIs and structured patient and clinicopathological data. Task 3, here- after Task Progression, evaluated patient-level time-to-progression prediction using H&E WSIs, RNA-seq measurements, and structured data. Both tasks used multi-institutional data, hidden validation and test s