A domain-adaptive deep contrastive network for magnetic resonance imaging-driven bladder cancer classification

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Oncology
• Urologic Oncology

Background:

• Bladder cancer presents significant morbidity and mortality.
• Deep learning shows potential for automated bladder cancer classification via MRI.
• Clinical use of deep learning is hindered by data discrepancies and poor feature distinction between non-muscle-invasive bladder cancer (NMIBC) and muscle-invasive bladder cancer (MIBC).

Purpose of the Study:

• To develop a novel deep learning framework, the Domain-Adaptive Deep Contrastive Network (DADCNet), for robust MRI-based bladder cancer classification.
• To improve cross-center generalization by learning domain-invariant representations.
• To enhance classification performance through improved feature discriminability.

Main Methods:

• Proposed a Domain-Adaptive Deep Contrastive Network (DADCNet) integrating source and target domain data for feature learning.
• Employed a deep contrastive learning strategy to boost inter-class separability and intra-class compactness.
• Validated the framework on a multi-center bladder cancer MRI dataset.

Main Results:

• DADCNet achieved superior performance compared to existing CNN and Transformer-based methods.
• The model attained an accuracy of 0.955, an F1-score of 0.955, and an AUC of 0.991.
• Demonstrated improved cross-center generalization and feature discriminability.

Conclusions:

• DADCNet effectively addresses challenges in multi-center MRI bladder cancer classification.
• The proposed domain-adaptive and contrastive learning approach leads to more robust and accurate classification.
• DADCNet shows significant promise for clinical deployment in bladder cancer diagnosis.