High-stakes staging problem. Bladder cancer is a urogenital cancer with high morbidity and mortality. Lymph node metastasis (LNM) is an independent risk factor for death, with the 5-year cancer-specific survival rate dropping to 27.7% for patients with LNM compared to much higher rates in those without.
Accurate preoperative prediction of lymph node status is essential for disease staging, selecting appropriate therapy, and estimating patient survival. Patients with confirmed LNM require perioperative systemic chemotherapy combined with extended lymph node dissection, making preoperative knowledge critical to treatment planning.
Conventional imaging methods including ultrasonography, CT, and MRI have poor sensitivity for detecting metastatic lymph nodes, with CT-based detection achieving only 31 to 45% sensitivity when relying on lymph node size criteria. Enlarged nodes may be reactive rather than malignant, while small nodes can harbor micrometastases, making size-based assessment fundamentally unreliable.
Extracting hidden image data. Radiomics is a high-throughput computational method that extracts large numbers of quantitative imaging features from medical images, enabling non-invasive analysis of the entire tumor including features invisible to the human eye.
Deep learning complements traditional radiomics by automatically learning feature representations from images without requiring manual feature engineering. Combining hand-crafted radiomics (HCR) features with deep learning (DL) features has shown promise for lymph node metastasis prediction in other cancers such as gastric and breast cancer.
Prior radiomics approaches for bladder cancer lymph node analysis faced a limitation: both positive and negative lymph nodes are typically small, meaning partial volume effects could reduce quantification accuracy. The authors focused instead on extracting features from the primary bladder tumor using three-phase CT to indirectly predict lymph node status.
Two-center retrospective design. The study enrolled 239 bladder cancer patients from two institutions: 185 in the training set from Qingdao University Hospital and 54 in the external test set from Shandong Provincial Hospital, spanning cases from March 2008 to June 2022.
All patients underwent three-phase pelvic CT within 20 days before surgery and received extended pelvic lymph node dissection up to the aortic bifurcation. Patients who received preoperative chemotherapy, radiotherapy, or immunotherapy were excluded to prevent confounding of the imaging features.
Three-phase CT captured images at corticomedullary phase (25 seconds), nephrographic phase (75 seconds), and excretory phase (300 seconds) after contrast injection, providing different tissue enhancement profiles across all three time points for comprehensive tumor characterization.
Clinical information collected included patient demographics, tumor location, shape, size, calcification, cystic necrosis, tumor boundary clarity, number of lesions, stalk presence, CT-reported T stage and lymph node status, and CT attenuation values in each of the three phases.
Combined radiomics and deep learning features. Manual tumor segmentation was performed in ITK-SNAP software, with intra- and inter-observer reliability assessed using intraclass correlation coefficients, retaining only features with ICC greater than 0.8 for reproducibility.
From each tumor region of interest, 2,645 hand-crafted radiomics features were extracted. A ResNet18 convolutional neural network pre-trained on a transfer learning platform extracted an additional 384 deep learning features, yielding a combined feature pool of 3,029 features per patient.
Feature selection proceeded in two steps: the Minimum Redundancy Maximum Relevance (mRMR) algorithm first retained 50 high-relevance, low-redundancy features, then LASSO logistic regression further reduced these to 12 features (8 hand-crafted and 4 deep learning) for building the final radiomics signature.
Combat compensation methodology was applied to eliminate scanner and protocol differences between the two imaging centers while retaining meaningful texture pattern characteristics, an important step for building generalizable multi-center models.
Nine classifiers tested systematically. Nine different machine learning classifiers were trained on the radiomics signature: support vector machine, logistic regression, XGBoost, NaiveBayes, AdaBoost, LightGBM, k-nearest neighbor, multilayer perceptron, and GradientBoosting.
Because approximately one-fifth of patients had lymph node metastasis, the dataset was imbalanced. The synthetic minority oversampling technique (SMOTE) was applied to retrain each classifier, generating synthetic minority-class samples to improve sensitivity for detecting the less common LNM-positive cases.
The best-performing radiomics signature was combined with significant clinical predictors identified through univariate and multivariate logistic regression to construct the final combined model, which was visualized as a clinician-friendly nomogram for practical use at the bedside.
Model performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, calibration curves to assess prediction agreement, and decision curve analysis to evaluate net clinical benefit across different risk thresholds. Kaplan-Meier survival curves assessed the model's prognostic risk stratification ability.
Key clinical risk factors identified. Univariate logistic regression identified stalk presence, CT-reported T stage, and CT-reported lymph node status as significant predictors of lymph node metastasis. Multivariate analysis confirmed that stalk presence and CT-reported LN status were independent predictors.
The clinical model alone achieved AUC values of 0.764 (95% CI: 0.697 to 0.831) in the training set and only 0.624 (95% CI: 0.402 to 0.846) in the external test set, reflecting the limited information captured by visual CT features and low-dimensional clinical parameters alone.
The LightGBM classifier retrained with SMOTE achieved the best radiomics signature performance, with an AUC of 0.893 (95% CI: 0.769 to 1.000) on the external test set. Among all features, wavelet-based texture uniformity demonstrated the greatest single-feature contribution to the model's predictive ability.
Best overall clinical performance. The combined model integrating the LightGBM-SMOTE radiomics signature with clinical predictors achieved an AUC of 0.834 (95% CI: 0.659 to 1.000) on the external test set, with accuracy of 0.870, the highest accuracy across all models tested.
Calibration curves demonstrated good agreement between predicted probabilities and actual LNM rates in both training and external test sets. Decision curve analysis showed the combined model provided superior net clinical benefit compared to the radiomics signature alone across clinically relevant threshold ranges.
Kaplan-Meier survival analysis confirmed that the combined model successfully stratified patients into high-risk and low-risk groups for progression-free survival in the overall cohort and training set, with statistically significant differences in survival curves (log rank p less than 0.05).
Risk stratification in the external test set did not reach statistical significance, likely due to selection bias: while fewer than 8% of LNM-positive patients in the training set survived beyond 30 months, 40% of LNM-positive patients in the external test set had progression-free survival exceeding 30 months.
Non-invasive staging tool. The combined CT-based deep learning radiomics model provides a non-invasive, readily available preoperative tool for predicting lymph node status in bladder cancer patients, potentially guiding treatment selection without requiring invasive lymph node sampling.
Activation maps from the convolutional neural network revealed that the highlighted regions associated with lymph node metastasis predictions corresponded to identifiable areas within the primary bladder lesion, providing interpretable evidence that the model captures biologically meaningful imaging patterns.
The study acknowledges key limitations including retrospective design, manual segmentation of tumor regions, and restriction to CT images alone. Future work aims to incorporate automatic segmentation, expand to multicenter international datasets, and add multimodal data including MRI, ultrasound, genomics, and pathology to improve generalizability.
The authors recommend including this CT-based combined model in bladder cancer predictive frameworks to support improved patient monitoring, treatment personalization, and adjuvant clinical trial design, helping to narrow the gap between radiology and precision oncology care.