.jpg)
Accurate breast cancer classification from mammograms requires effective integration of craniocaudal (CC) and mediolateral oblique (MLO) views, as each provides complementary information about anatomy and potential abnormalities. However, traditional multi-view learning approaches often limit themselves to aggregating features in a single layer or employing cross-attention mechanisms at a single stage, which tends to intermix view-specific representations and restricts interaction to shallow levels of the neural network. Faced with this limitation, a promising paradigm emerges: token-based dual-view fusion, a conceptual framework that reformulates inter-view interaction as structured token-level communication within a pre-trained and frozen transformer.
In this approach, instead of directly merging feature maps, dedicated fusion tokens are introduced that act as intermediate carriers of cross-dependencies between the CC and MLO views. These tokens exchange information bidirectionally through cross-attention mechanisms and are inserted at multiple depths of the transformer encoder, enabling progressive and hierarchical interaction throughout the entire architecture. In this way, the fusion tokens are reintegrated into the token sequence and refined by subsequent layers, while preserving the specific structure of each view. This design avoids representation entanglement and maximizes information complementarity.
Experimental results on datasets such as VinDr-Mammo and CMMD demonstrate consistent improvements over linear tuning methods, prompt-only adaptation, and conventional fusion baselines. In the BI-RADS classification task, an F1-score of 50.40% and an AUC of 0.8090 are achieved, with a 0.10 increase in AUC over the dual fusion baseline in the binary scenario. These advances open the door to more accurate and reliable computer-aided diagnosis systems, especially when integrated into custom software platforms for clinical environments.
Implementing these advanced computer vision techniques requires a solid artificial intelligence foundation and a robust technological ecosystem. At Q2BSTUDIO, as specialists in artificial intelligence for businesses, we develop custom applications incorporating deep learning models, AI agents, and medical image processing solutions. Additionally, we combine these developments with AWS and Azure cloud services to ensure scalability, cybersecurity in handling sensitive data, and seamless integration with business intelligence tools such as Power BI. Our approach enables healthcare organizations to fully leverage innovations in breast cancer classification, transforming academic research into operational, high-value clinical solutions.
