An Efficient Cross-Modal Interaction and Dynamic Fusion Network for Multimodal Breast Ultrasound Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Multimodal breast ultrasound (B-mode, Doppler, elastography) offers complementary data for lesion characterization.
• Integrating these heterogeneous modalities is challenging due to inconsistent features, limited interaction, computational costs, and data noise/missingness.

Purpose of the Study:

• To develop an efficient Cross-Modal Interaction and Dynamic Fusion Network (CIDFNet) for enhanced multimodal breast ultrasound analysis.
• To address challenges in feature integration, cross-modal interaction, and data reconstruction for improved lesion characterization.

Main Methods:

• Proposed CIDFNet integrates multi-scale feature enhancement, early-stage cross-modal interaction, and dynamic fusion based on feature reliability.
• An invertible neural network was used for reconstructing missing modality features during training.
• The framework aims for a balance between performance and computational efficiency.

Main Results:

• CIDFNet achieved an AUC of 85.69%, accuracy of 75.51%, recall of 50.00%, F1-score of 62.50%, and precision of 83.33% on an internal dataset.
• The model requires 49.51 M parameters and 79.79 G FLOPs.
• Performance degradation was observed under Gaussian noise perturbation.

Conclusions:

• CIDFNet provides an effective framework for multimodal breast ultrasound analysis.
• The network demonstrates a practical trade-off between diagnostic performance and computational efficiency.
• Further research may explore robustness to various noise conditions.