Diagnosis depends on human inspection. Bladder cancer (BCa) is the fourth most common malignant tumor in men, and accurate diagnosis currently relies on histopathology performed by pathologists examining tissue slides under a microscope by the naked eye.
Challenging edge cases exist. Some histopathology patterns, such as microcystic urothelial carcinoma and papillary urothelial neoplasm of low malignant potential, can have deceptive appearances that even immunochemistry methods may not resolve reliably.
Machine learning offers a new path. High-throughput processing of medical images has shown promise across multiple cancer types, but its application to pathology images in bladder cancer had not been fully explored, motivating this study to build and verify automated diagnostic and prognostic models.
Two independent patient cohorts were assembled. The General cohort comprised 108 patients who underwent radical or partial cystectomy at Shanghai General Hospital between 2009 and 2016. A TCGA cohort of 406 patients with available digital pathology images was also enrolled.
Tissue preparation followed a rigorous protocol. FFPE samples were sliced at 5 micrometers, stained with hematoxylin and eosin, and reviewed by an experienced genitourinary pathologist. Representative 1000 x 1000 pixel images were captured at 400x magnification, targeting features such as nuclear pleomorphism, mitosis, and tumor invasion.
Normal and benign tissues were included as controls. In addition to cancer samples, 53 normal bladder and 39 glandular cystitis FFPE samples from Shanghai General Hospital were processed identically, enabling the model to distinguish cancer from both normal tissue and a clinically similar benign condition.
A total of 643 images were produced. All representative images were independently reviewed by a second pathologist to ensure academic rigor before being used in model training and validation.
CellProfiler was used to extract 345 quantitative features. An image processing pipeline segmented cell nuclei and cytoplasm, then measured object shape, size, texture, and pixel intensity distribution from each H&E image.
LASSO analysis selected the most informative features. Least Absolute Shrinkage and Selection Operator (LASSO) with 10-fold cross-validation was applied to identify the most relevant image features and assign weighted coefficients, producing a compact and interpretable diagnostic score formula.
A separate survival model was built using LASSO-Cox analysis. For the prognosis model, LASSO-Cox regression was applied to the TCGA training cohort to identify survival-related image features, and a machine learning-based risk score was derived from their weighted sum.
An integrated nomogram combined AI scores with clinical data. The final prognostic tool synthesized the machine learning risk score with clinicopathologic factors such as tumor stage, grade, and patient age, evaluated by calibration with bootstrapping plots.
22 image features were identified for cancer detection. LASSO analysis with 10-fold cross-validation selected 22 BCa-related image factors that formed the basis of the diagnostic model.
AUC values exceeded 89% in every tested cohort. The model achieved AUC values of 96.3% in the training cohort, 89.2% in the test cohort, and 94.1% in the external validation cohort when distinguishing BCa from normal bladder tissue.
The model also distinguished BCa from glandular cystitis. This is a clinically important differentiation because glandular cystitis can appear similar to cancer. The model achieved an AUC of 93.4% in that comparison, and 93.8% when BCa was compared against all non-cancer samples combined.
18 survival-related image features were identified. LASSO-Cox analysis of the TCGA training cohort selected 18 image factors associated with overall survival in bladder cancer patients.
High risk scores correlated with advanced stage and grade. Patients with higher machine learning risk scores were significantly more likely to have high-stage (T3 or T4) and high-grade tumors in both the TCGA and General cohorts.
Survival differences were large and statistically significant. High-risk versus low-risk patients had a hazard ratio of 2.09 in the TCGA cohort and 5.32 in the General cohort, both with p-values below 0.0001, demonstrating a strong and consistent survival stratification effect.
The risk score was an independent prognostic factor. Univariate and multivariate Cox regression confirmed the risk score acted independently of other clinical variables in predicting survival in both cohorts, giving it real standalone prognostic value.
An integrated nomogram was built to improve prediction. By combining the machine learning risk score with conventional clinical factors including tumor stage, grade, and patient age, researchers created a nomogram designed to surpass standard TNM-based survival prediction.
Calibration plots confirmed accuracy. Bootstrapping calibration plots showed strong agreement between predicted and observed survival at 1, 3, and 5 years, supporting the reliability of the nomogram in real-world applications.
AUC values were higher than those from clinical factors alone. ROC curves for 1-, 3-, and 5-year survival yielded AUC values of 77.7%, 83.8%, and 81.3% respectively in the TCGA cohort. These values were consistently higher than predictions based only on stage or grade.
Results were confirmed in the General cohort. Incremental survival prediction gains from the nomogram were also observed in the independent General cohort, demonstrating the stability and reproducibility of this integrated approach.
Pathomics directly digitizes tumor tissue. Unlike radiomics, which analyzes radiation-based images that indirectly reflect tumor biology, pathomics scans actual histological slides, potentially capturing richer information about tumor cell characteristics and the extracellular matrix.
This study's approach is clinician-friendly. The automatic image processing pipeline and LASSO-based model were designed so that clinicians without computational expertise could use the resulting scores in practice, lowering the barrier to real-world adoption.
Stage T2 and T3 tumors are particularly difficult to assess. The limitations of microscopy-based biopsy at these stages make an AI-assisted tool especially valuable for improving diagnostic confidence in cases where traditional methods struggle.
The retrospective design limits immediate clinical translation. The study was retrospective, and the authors acknowledge that prospective trials designed according to SPIRIT-AI and CONSORT-AI guidelines are needed before the model can be adopted in clinical workflows.
Risk score cut-off varied between cohorts. The cut-off separating high and low risk was defined by the median score within each cohort separately, meaning the absolute threshold differs between TCGA and General cohort patients, which complicates direct comparison.
The model did not fully match expert pathologist accuracy. In certain comparisons, the machine learning model was less accurate than traditional histopathology performed by trained pathologists, identifying an area requiring further refinement.
Despite limitations, the approach is genuinely novel. This study was among the first to apply pathological image segmentation to automated BCa diagnosis and prognosis, introducing a new type of prognostic indicator distinct from conventional clinicopathologic factors.