Bladder cancer staging determines whether patients need radical surgery or organ-sparing treatment. Muscle-invasive bladder cancer (MIBC) requires radical cystectomy, while non-muscle-invasive bladder cancer (non-MIBC) allows bladder preservation. Accurate preoperative distinction between these two categories is therefore critical for treatment planning.
Current standard diagnostic methods have significant limitations. Cystoscopy with biopsy is the gold standard but can cause urethral injury, urinary infection, and undersampling of the tumor. Some patients diagnosed as T1 stage (non-invasive) during transurethral resection are found to have muscle invasion at subsequent radical cystectomy, leading to inadequate initial treatment.
CT imaging alone achieves only 35 to 40 percent accuracy for T-staging bladder cancer. MRI offers better soft tissue contrast but is limited by high cost, long scanning duration, and contraindications. A more accurate non-invasive preoperative staging tool is needed to reduce under-treatment and over-treatment.
Radiomics and deep learning offer complementary advantages for image-based cancer staging. Radiomics extracts quantitative imaging biomarkers from tumors that are invisible to the human eye. Deep learning using convolutional neural networks can automatically learn high-dimensional spatial and textural features from CT images without requiring manual feature engineering.
The study retrospectively analyzed 173 bladder cancer patients treated from 2017 to 2021. Of these, 43 had pathologically confirmed MIBC and 130 had non-MIBC, confirmed by histopathology from radical or partial cystectomy or transurethral resection. CT scans were performed within four weeks of surgery.
Patients were split 75 to 25 percent into training and test cohorts. A total of 129 patients formed the training set and 44 the test set. The two cohorts were balanced across clinical characteristics including age, number of lesions, hydronephrosis status, gender, and therapeutic approach, with no statistically significant differences.
CT imaging was performed on two standardized scanner platforms. Philips ICT 256 and Philips IQon spectral CT were used at 120 to 130 kV with 1-mm slice thickness. Three contrast phases were acquired at 30, 60, and 300 seconds after contrast injection to capture corticomedullary, nephrographic, and excretory phases.
Tumor segmentation was performed manually by two radiologists to assess reproducibility. The entire tumor volume was delineated using ITK-SNAP software. When multiple tumors were present, the largest lesion was selected for feature extraction. Intraclass correlation coefficients greater than 0.80 confirmed good inter-observer agreement.
A total of 1,735 radiomics features were extracted from CT images using PyRadiomics. These included 18 geometric features, 14 intensity features, and 75 texture features, further expanded by applying logarithmic, wavelet, gradient, and local binary pattern transforms. Images were resampled to 1x1x1 mm voxels before extraction.
Features were filtered for reproducibility and redundancy before model training. Only features with intraclass correlation coefficients above 0.80 were retained, yielding 1,616 stable radiomics features. Spearman rank correlation identified and removed highly correlated redundant features using a greedy recursive strategy.
Deep learning features were extracted using a pretrained ResNet50 convolutional neural network. The largest tumor slice was resized to 224 by 224 pixels and passed through ResNet50, originally trained on the ImageNet dataset. This produced 2,048 deep features per patient, which were then compressed to 128 dimensions using Principal Component Analysis.
Combined deep learning and radiomics features were further reduced by LASSO regression. After fusing the 1,735 radiomics features with the 128 deep transfer learning features, LASSO with 10-fold cross-validation selected 30 final features: 11 radiomics and 19 deep learning features. These formed the deep learning radiomics signature used for all downstream models.
Eight machine learning classifiers were trained and compared on the deep learning radiomics features. Models tested included logistic regression, SVM, KNN, Random Forest, ExtraTrees, XGBoost, LightGBM, and a multilayer perceptron (MLP). All hyperparameters were optimized using grid search with cross-validation via the scikit-learn library.
The MLP classifier used three hidden layers with progressively smaller dimensions. Hidden layer sizes of 256, 128, and 64 neurons were configured, optimized with the Adam optimizer at a learning rate of 0.0001. This architecture was chosen for its capacity to learn nonlinear interactions among the combined radiomics and deep learning features.
A cosine decay learning rate schedule was applied during ResNet50 training to improve generalization. The learning rate oscillated between 0 and 0.01 across 30 training epochs. Stochastic gradient descent with sigmoid cross-entropy loss was used to optimize the deep feature extractor before feature extraction.
Model performance was evaluated using AUC, accuracy, sensitivity, specificity, PPV, and NPV. ROC curves were constructed for each classifier in both the training and test cohorts. Decision curve analysis was used to assess the net clinical benefit of the models across a range of decision thresholds.
The MLP model achieved the best test cohort AUC of 0.884 among all classifiers. Training AUC was 0.973 (95% CI 0.949 to 0.998), and test AUC was 0.884 (95% CI 0.783 to 0.986). In the test cohort, the MLP achieved sensitivity of 91 percent, specificity of 79 percent, and accuracy of 82 percent.
The combined deep learning radiomics model outperformed radiomics-only and deep learning-only models. The radiomics-only MLP achieved a test AUC of only 0.78, while the deep transfer learning model alone achieved 0.76 in the test cohort. Fusing both feature types into the deep learning radiomics signature yielded the highest diagnostic accuracy.
Other classifiers showed strong training performance but limited test set generalization. SVM, Random Forest, and ExtraTrees models achieved near-perfect training AUCs above 0.99 but dropped substantially in the test cohort, suggesting overfitting. The MLP showed the most stable performance gap between training and testing, demonstrating better generalization.
Decision curve analysis confirmed the clinical utility of the MLP-based deep learning radiomics model. Across most decision thresholds, the combined model provided greater net benefit than treating all patients or no patients, outperforming the radiomics-only model. This validates the practical value of adding deep learning features beyond what radiomics alone provides.
ResNet50 was selected as the deep feature extractor for its residual block architecture. Unlike standard CNNs, ResNet's skip connections allow training much deeper networks without gradient degradation, enabling the model to capture subtle textural and structural differences between MIBC and non-MIBC tumors in CT images.
Transfer learning from ImageNet weights was used to compensate for limited medical training data. Medical imaging datasets are small compared to what neural networks typically require. Initializing ResNet50 with ImageNet weights allows it to transfer general visual feature extraction capabilities to the bladder cancer CT domain with minimal overfitting risk.
PCA dimensionality reduction was essential for processing the 2,048 ResNet50 output features. Without dimensionality reduction, the high-dimensional deep feature space would overwhelm downstream classifiers and increase overfitting risk. Compressing to 128 principal components preserved the most informative variance while enabling stable classification.
The deep learning features provided complementary information to radiomics. While radiomics quantifies hand-crafted statistical properties of tumor regions, deep learning extracts high-dimensional representations that may capture tissue heterogeneity and spatial patterns invisible to conventional feature engineering. Their combination consistently improved performance over either approach alone.
The single-center retrospective design limits generalizability of the findings. All patients were treated at Zhongshan People's Hospital and CT images were acquired on two scanner models. Multi-center prospective validation with external datasets is needed to confirm the model's performance in varied clinical settings.
Class imbalance between MIBC and non-MIBC may have reduced radiomics-only model performance. With 43 MIBC versus 130 non-MIBC patients, the training set contained three times as many non-invasive cases. This imbalance likely contributed to lower sensitivity and AUC in classifiers relying only on hand-crafted radiomic features.
Only the arterial phase CT images were used, limiting access to potentially informative later-phase data. Venous and excretory phase images capture different aspects of tumor vascularity and urinary tract anatomy. Including all three contrast phases in future models may further improve staging accuracy.
Automated segmentation methods should replace manual ROI delineation in future work. Manual tumor segmentation by radiologists is time-consuming and subject to variability, even when reproducibility is good by ICC standards. Deep learning-based semantic segmentation models could automate this step and increase throughput and reproducibility.
The transfer learning approach using ImageNet weights may be suboptimal for medical image features. Models pretrained on natural images learn features specific to photographic scenes rather than tissue pathology. Pre-training on large medical imaging datasets in future studies could improve feature relevance and model accuracy.
This study presents the first application of deep learning radiomics for preoperative bladder cancer staging from CT images. The combined approach achieved clinically meaningful accuracy in distinguishing MIBC from non-MIBC without requiring MRI or invasive biopsy, offering a non-invasive preoperative staging option.
Accurate preoperative staging could reduce the rate of under-staged patients receiving inadequate treatment. Patients incorrectly classified as non-MIBC who undergo TURBT alone when radical cystectomy is needed face higher recurrence risk. An accurate CT-based classifier could flag these patients for more aggressive staging workup or direct cystectomy.
The model's high negative predictive value of 96 percent in the test cohort is clinically valuable. A high NPV means that patients predicted to have non-MIBC are very unlikely to actually have muscle invasion, supporting conservative bladder-sparing decisions with greater confidence.
Integration of multi-phase CT and automated segmentation may enable clinical deployment. With further refinement and validation, a deep learning radiomics staging tool could be embedded into CT reporting workflows to assist radiologists and urologists in making more accurate treatment decisions before surgery.