Tumour budding refers to discrete clusters of up to four cancer cells found at the invasive margins of a tumour, and it has been identified as a prognostic marker in pancreatic, colorectal, breast, and lung cancers -- but had never been studied in muscle-invasive bladder cancer before this work. These small cell clusters are thought to represent an early step toward metastasis, reflecting epithelial-to-mesenchymal transition in the migrating cancer cells.
Muscle-invasive bladder cancer is classified as tumours that have grown into or through the bladder muscle layer and carries a five-year survival rate of 30 to 50%. Despite three decades of research, mortality rates have remained stubbornly unchanged, motivating the search for prognostic features beyond the current TNM staging system.
TNM staging captures anatomic extent of disease but does not reflect the biological heterogeneity within each stage. Two patients with identical TNM stage can have dramatically different outcomes, suggesting that additional morphological or molecular markers are needed to refine risk stratification.
Manual tumour bud counting is impractical for routine clinical use due to observer variability and the time required to count buds across large tissue sections. An automated machine learning approach that quantifies buds consistently across whole slide images would enable systematic study and eventual clinical adoption.
One hundred patients who underwent radical cystectomy for muscle-invasive bladder cancer at Edinburgh hospitals between 2006 and 2013 were enrolled. The cohort started at 163 patients but was reduced to 110 due to unavailable tissue, and a further 10 were excluded for unknown death cause or extensive tissue artefacts, yielding a final study group of 100 patients.
The cohort was 59% male with a median age of 68 years. In terms of TNM staging, 25% had stage II disease, 45% had stage III (41% IIIA, 4% IIIB), and 30% had stage IV. Follow-up extended to 113 months with a median disease-specific survival of 23.63 months. Fifty-four patients died of bladder cancer during follow-up.
Disease-specific survival was the primary endpoint, defined as death directly attributable to bladder cancer rather than other causes. Data were collected from NHS Scotland's patient database, providing objective outcome ascertainment independent of the research team.
The archived formalin-fixed paraffin-embedded tissue block containing the deepest cancer invasion was selected for each patient after review by both a pathologist and research scientist. This selection ensured that the most clinically informative tissue was analyzed for tumour budding quantification.
The automated pipeline used immunofluorescence staining with a Pan-cytokeratin antibody to visualize urothelial cells against a Hoechst nuclear counterstain, then applied two complementary deep learning models to segment tumour regions from stroma and artefacts. Using fluorescence rather than standard H and E staining allowed clearer identification of tumour buds even in areas with dense inflammatory infiltrate.
Tumour region segmentation was performed by ensemble prediction from two neural network models. The first was a semantic segmentation convolutional neural network based on the Inception architecture with transposed convolution layers and skip connections. The second combined layer normalization, a fully convolutional network, and a classification random forest. The outputs of both models were merged to produce a final tumour mask, achieving a Dice score of 0.86 against manual annotations.
Nuclei were detected separately using two regression visual context random forest models that predicted spatial proximity to nuclei and nucleus size. A local maxima algorithm then derived nucleus positions from these prediction maps. Pearson correlation between automatically detected and manually annotated nucleus counts per tumour cell cluster was 99.3%, confirming high detection accuracy.
Tumour buds were defined as connected tumour cell clusters containing between one and four nuclei, consistent with international colorectal cancer reporting recommendations adapted for bladder cancer. Seven quantification features were computed per patient, covering number and density of buds in both the tumour core and the invasive front, across four different counting area definitions.
Among all seven tumour bud features, the total number of tumour buds in the detected tumour core was the most strongly associated with disease-specific survival. Its hazard ratio by Cox regression was 3.22 (p = 2.67E-05), making it the second most prognostic feature in the entire dataset after TNM stage (HR = 5.44, p = 2.53E-09).
High tumour bud count in the core was correlated with higher TNM stage (chi-squared p = 0.00089) and higher pT stage (p = 0.0078), the two most prognostically important clinical features. However, tumour budding provided independent prognostic information beyond what these staging variables captured.
Intra-tumoural budding in the tumour core outperformed peri-tumoural budding at the invasive front for survival prediction. This contrasts with findings in colorectal cancer where peri-tumoural budding is typically the dominant signal, suggesting that the spatial pattern of budding may differ between cancer types.
Tumour grade and growth pattern were not prognostic in this cohort despite being standard pathologic descriptors. Only TNM staging, pT stage, lymph node status, and core tumour budding showed statistically significant associations with disease-specific survival in univariate analysis.
A survival decision tree combining clinical and tumour budding features stratified patients into three groups more precisely than standard TNM staging. The tree first split all 100 patients using TNM stage, separating stage IV metastatic disease from stage II and III. Within the non-metastatic group, the number of tumour buds in the core was the most significant feature for the second split, chosen in 97 of 100 leave-one-out cross-validation runs.
The resulting TB stage model reassigned 27 patients previously classified as TNM stage III into a lower-risk stage II' group. This reclassification improved separation between surviving and non-surviving patients, as these 27 patients had survival characteristics more similar to stage II than to the remaining stage III patients.
The TB stage model produced statistically significant separation between all three groups (p less than 0.05 for both stage III' and stage IV versus stage II'). In contrast, standard TNM staging failed to achieve significant separation between stage II and stage III patients (p = 0.13), with the confidence interval for stage III hazard ratio including 1.0.
Overall model fit statistics including the likelihood ratio test, Wald test, and log rank test all yielded larger test statistics for the TB stage model than for clinical TNM staging, indicating stronger rejection of the null hypothesis that stage has no effect on survival.
The deep learning algorithms were designed to be adaptable from immunofluorescence to standard H and E or immunohistochemistry staining used in routine pathology labs. Because the models can be retrained on brightfield microscopy annotations, the methodology does not require fluorescence scanning equipment, which is not available at most clinical sites.
Automated quantification addresses the main barriers to clinical adoption of tumour budding assessment: inter-observer variability among pathologists and the time required for manual counting across whole slides. The machine learning pipeline eliminates both sources of inconsistency while enabling analysis of the entire tissue section rather than selected fields of view.
Whole-slide quantification methods evaluating the complete tumour core outperformed traditional field-of-view methods that count buds within small representative areas, suggesting that global tumour bud burden carries more prognostic information than local hotspot counts.
The study also identified that the same nuclei detection and tumour segmentation infrastructure can be extended to immune cell quantification, potentially enabling a combined tumour budding and immune profiling analysis that further improves prognostic stratification beyond what either feature provides alone.