Staging system falls short for aggressive bladder cancer. Muscle-invasive bladder cancer (MIBC) accounts for the majority of bladder cancer mortality, with more than 50% of patients dying from metastatic disease within 5 years despite radical cystectomy being the gold-standard treatment.
The current TNM staging system describes anatomical depth of invasion and node status, but is purely anatomy-based. Patients with identical TNM stages frequently experience vastly different outcomes, reflecting the system's inability to capture the biological aggressiveness of individual tumors.
The tumor microenvironment plays a critical role in cancer progression. The location, density, functional state, and organization of immune cells within the tumor, collectively termed the immune contexture, have emerged as potentially more informative prognostic factors than anatomical staging alone, but are not yet included in clinical TNM guidelines.
Four key immune biomarkers studied. The study focused on four interconnected components of the tumor immune microenvironment: tumor-infiltrating lymphocytes (TILs), tumor-associated macrophages (TAMs), tumor buds (TBs), and the immune checkpoint ligand PD-L1.
Cytotoxic CD8 T-cells are the primary anti-tumor effectors, and their presence in the tumor microenvironment is consistently associated with better survival in many cancer types. Macrophages can be polarized toward anti-tumor M1 or pro-tumor M2 phenotypes, with M2 macrophages promoting tumor cell dissemination and disrupting T-cell function during metastasis.
Tumor budding, defined as isolated single cancer cells or clusters of up to four cells at the invasive front, represents the first step of cancer metastasis and has been added to TNM staging as a supplementary factor for colorectal cancer. PD-L1 expression by tumor cells suppresses TIL activity by binding to PD-1 receptors, enabling immune evasion and tumor progression.
First whole-slide multiplex labeling of MIBC. Tissue specimens from 78 MIBC patients who underwent radical cystectomy at Edinburgh hospitals between 2006 and 2013 were processed using automated tyramide-based multiplex immunofluorescence, marking the first study to label entire MIBC tissue sections with multiple simultaneous fluorescence immune markers.
Six primary antibodies were applied simultaneously: Pan-cytokeratin (cancer cells), CD3 (general T-cells), CD8 (cytotoxic T-cells), CD68 (total macrophages), CD163 (M2 macrophages), and PD-L1 (immune checkpoint ligand). This multiplex approach allows simultaneous visualization of multiple cell populations that single-marker immunohistochemistry cannot achieve.
Stained slides were scanned at 20x magnification into whole-slide fluorescence images. Features were quantified across two anatomically defined tumor regions: the tumor core (main tumor mass) and the invasive front (a 1000 micrometer border around the tumor core, split equally inside and outside the tumor edge), recognizing that immune cell distribution varies meaningfully between these regions.
Machine learning-driven cell detection. A multi-stage random forest pipeline detected and localized all cell nuclei across whole-slide images using classification, proximity map, and surface area map models trained on expert annotations. Each detected cell was then classified by its immunofluorescence signal intensity for each marker.
Tumor buds were quantified by a CNN-RF hybrid model: a convolutional neural network generated coarse epithelium segmentation masks, which were then refined by a random forest. Clusters of 1 to 4 epithelial nuclei were classified as tumor buds, enabling reproducible automated quantification that overcomes the poor inter-observer consistency of manual budding assessment.
Beyond cell counts and densities, pairwise spatial relationships were quantified using Ripley's L function, which measures whether one cell population is dispersed, randomly distributed, or clustered around another at distances ranging from 20 to 250 micrometers. This captured the proximity of immune cells to tumor buds and to each other at fine spatial scales.
In total, 201 features were extracted per patient: 126 image features (cell counts, densities, PD-L1 expression), 60 spatial features (pairwise cell distribution patterns), and 15 clinical features (age, sex, TNM stage). These were organized into seven feature combinations to test which combination best predicted 5-year survival.
Rigorous nested cross-validation approach. Five ML algorithms with fundamentally different theoretical approaches were evaluated: decision tree, random forest, support vector machine, linear regression, and k-nearest neighbors. Nested cross-validation with separate training and test sets was used to prevent overfitting in this small 78-patient dataset.
Hyperparameter tuning used random search over 200 configurations per algorithm, shown to outperform grid search in high-dimensional spaces. Algorithm selection used two-fold outer cross-validation after five-fold inner cross-validation for hyperparameter optimization, ensuring each algorithm was evaluated on multiple held-out data splits.
Rather than selecting a single best classifier, the final ensemble model combined four complementary submodels: a linear SVM using image features, a decision tree using image and clinical features, a logistic regression using image and spatial features, and a random forest using all features. A patient was classified as high-risk if two or more submodels predicted poor prognosis.
Dramatic improvement over current gold standard. On the independent test set, the ensemble model achieved 89.3% AUROC compared to 64.3% for TNM staging. Accuracy was 80% versus 50% for TNM, and the F1 score was 83.3% versus 44.0%.
The most striking difference was in specificity for identifying patients who would die: the ensemble model correctly classified 71.4% of patients who succumbed to MIBC, compared to only 28.6% correctly classified by TNM staging. This near 2.5-fold improvement in identifying high-risk patients is clinically meaningful for treatment intensification decisions.
The ensemble model produced highly significant patient risk stratification (p value less than 0.00001) with a hazard ratio of 32.5 on the test set, compared to a hazard ratio of only 3.3 for TNM staging, demonstrating far greater separation between predicted low- and high-risk survival curves.
Tumor buds and immune cell densities drive predictions. Post-hoc analysis across all four submodels consistently identified tumor bud density in both the invasive front and tumor core as indicators of poor prognosis, reinforcing the clinical relevance of tumor budding as a marker of aggressive behavior in MIBC.
High densities of CD8+ cytotoxic T-cells, CD3+ general T-cells, and CD68+ macrophages in the invasive front and tumor core were consistently associated with good prognosis across multiple submodels, supporting the protective role of active immune infiltration in the tumor microenvironment.
Spatial features provided additional insight: CD3+ T-cells clustered within 20 micrometers of tumor buds were associated with good prognosis, while PD-L1 expression near tumor buds and M2 macrophages at certain distances contributed to risk predictions. The complex relationship between CD163+ M2 macrophages and PD-L1 expression produced context-dependent effects on prognosis across different submodels.
Case for immune contexture in clinical staging. The findings demonstrate that standardized quantification of immunological features from whole-slide images, combined with clinical data, can substantially improve bladder cancer prognosis prediction beyond what anatomy-based TNM staging provides.
Multiplex immunofluorescence enables simultaneous capture of the tumor-immune architecture from the cellular to subcellular level in a single tissue section, providing far more information about the microenvironment than standard H&E staining or single-marker immunohistochemistry approaches used in current clinical practice.
The authors argue for incorporating immune contexture assessment into clinical MIBC management, following the precedent of the Immunoscore system now validated for colorectal cancer. Identifying the 71.4% of high-risk MIBC patients who will not survive 5 years could enable earlier treatment intensification, improved clinical trial design, and more precise patient counseling.
Future work should validate these findings in larger multicenter cohorts and explore integration with genomic data to further refine patient stratification. The methodology's reliance on automated whole-slide analysis also supports scalability for clinical deployment in digital pathology workflows.