Magnetic resonance imaging plays an important role in bladder cancer staging, but distinguishing non-muscle-invasive from muscle-invasive disease remains difficult. Tumours can have irregular shapes, unclear boundaries and small lesion areas, while MRI data often vary across centres. DADCNet was developed to improve classification by addressing both feature discrimination and cross-centre variation.
A Framework Built for Variable MRI Data
DADCNet combines a convolutional neural network with domain adaptation and contrastive learning in one framework. The model is designed to learn from source-domain and target-domain MRI samples at the same time, so that it can generate feature representations that remain discriminative while becoming less sensitive to inter-centre variation. This directly addresses the problem of domain shift, which remains a major obstacle when models trained in one setting are applied to data from another.
Must Read: DL Model Predicts Lymphovascular Invasion in Bladder Cancer
The framework begins by extracting latent feature representations from MRI images through a CNN-based feature extraction module. Features from the source domain are then passed to a classifier that predicts whether a case belongs to the non-muscle-invasive or muscle-invasive category. At the same time, a domain adaptation module works to reduce the discrepancy between source-domain and target-domain features. Wasserstein distance is used to measure and minimise that discrepancy, helping align heterogeneous MRI data while preserving the information needed for classification.
A contrastive learning strategy forms the second main part of the design. It encourages images from the same class to move closer together in feature space while pushing images from different classes farther apart. This improves compactness within each class and separation between classes, making the distinction between non-muscle-invasive and muscle-invasive disease clearer. The model therefore aims not only to classify disease status, but also to remain robust when the imaging distribution changes across centres.
Better Performance Across Compared Methods
The model was evaluated on a multi-centre bladder cancer MRI dataset using cross-validation and was compared with several established CNN-based and Transformer-based methods. These included ResNet, EfficientNet, ViT, DeiT, Swin Transformer, MaxViT and ConViT. Across the reported comparisons, DADCNet achieved the strongest overall results on the main classification metrics. Its reported performance included an accuracy and F1-score of 0.955 and an AUC of 0.991.
Among the comparison models, ResNet and EfficientNet were the strongest conventional CNN baselines, while MaxViT was the strongest Transformer-based baseline. Even against these more competitive methods, DADCNet remained ahead. The gains over the stronger CNN baselines were modest but consistent, and the margins were wider against several Transformer-based models. This pattern indicates that the framework improved both classification performance and class discrimination rather than producing a narrow gain on a single metric.
Ablation analysis clarified the contribution of the added modules. When the combined contrastive learning and domain adaptation components were removed, performance decreased and training became less stable. The simplified version produced lower values for accuracy, F1-score and AUC than the full model. The full framework therefore benefited from both additions. Domain adaptation reduced inter-centre discrepancies, while contrastive learning sharpened class separation in the latent feature space. Together, the two components strengthened robustness and generalisation, which were central aims of the model design.
Generalisation and Interpretability Support Clinical Use
Cross-centre generalisation was assessed through several training and testing scenarios built from data collected at four medical centres. In each scenario, the model was trained on data from multiple centres and then tested on a different centre. Performance improved as training progressed and became more stable in later epochs, indicating that the framework adapted more effectively to heterogeneous data over time.
One scenario produced the strongest overall results, particularly in later training stages, suggesting that broader variation in the training data supported better generalisation. Confusion matrices also showed that misclassification between non-muscle-invasive and muscle-invasive cases decreased as training continued. Predictions became more balanced by the final epoch, supporting the view that the model learned increasingly reliable representations.
Interpretability analysis added a further layer of support. Grad-CAM visualisations showed that the model mainly focused on tumour regions and the adjacent muscular layer when distinguishing the two disease categories. These are the anatomical regions used in muscle invasion assessment, which supports the clinical relevance of the learned attention patterns. A separate t-SNE analysis showed that the feature distributions of the two classes overlapped at earlier stages of training but became more clearly separated as training progressed. The dataset used for evaluation comprised 279 patients with T2-weighted imaging sequences collected from four centres, providing a basis for testing the model under variable imaging conditions.
DADCNet was developed to address two persistent challenges in MRI-based bladder cancer classification: variation between centres and limited separation between non-muscle-invasive and muscle-invasive disease in feature space. By combining CNN-based feature extraction with domain adaptation and contrastive learning, the framework achieved stronger performance than the compared baseline methods and showed improved cross-centre generalisation. Interpretability analysis also indicated that the model focused on tumour regions and the adjacent muscular layer, supporting the clinical relevance of its predictions. Limitations remain, including the relatively limited dataset size, variations in MRI acquisition protocols across centres and the computational cost of model training. Even with these constraints, the reported findings indicate potential for supporting preoperative diagnosis and individualised treatment planning in bladder cancer.
Source: npj Digital Medicine
Image Credit: iStock
References:
Huang J, Hu H, Sun M et al. (2026) A domain-adaptive deep contrastive network for magnetic resonance imaging-driven bladder cancer classification. npj Digit Med: In Press.
