UltraMamba: Mamba-Based Multimodal Ultrasound Image Adaptive Fusion for Breast Lesion Segmentation

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Biomedical engineering

Background:

• Multimodal ultrasound imaging (B-mode, shear wave velocity, shear wave time) is vital for breast lesion diagnosis.
• Challenges include intermodal feature misalignment and attention shifts, hindering accurate segmentation.
• Existing methods often overemphasize color data, neglecting crucial textural and quantitative information.

Purpose of the Study:

• To introduce a novel segmentation framework, UltraMamba, and a comprehensive multimodal ultrasound breast lesion dataset (BreLS).
• To improve breast lesion segmentation accuracy by addressing intermodal feature misalignment and enhancing region-specific information.
• To provide a valuable resource for the analysis of multimodal ultrasound breast lesion data.

Main Methods:

• Developed the UltraMamba framework featuring bidirectional alignment and region-aware feature enhancement.
• Introduced the BreLS dataset, the first 2D multimodal ultrasound breast lesion dataset with 506 cases.
• Utilized Cross-Modal Knowledge Interaction and Region-Aware Feature Excitation modules within UltraMamba.

Main Results:

• UltraMamba achieved a Dice Similarity Coefficient (DSC) of 72.16% and HD95 of 42.02 mm on the BreLS dataset.
• Demonstrated a 2.59% improvement in DSC and a 6.78 mm reduction in HD95 compared to the MMCA-NET framework.
• The framework effectively improved segmentation accuracy, overcoming challenges of feature misalignment.

Conclusions:

• UltraMamba significantly enhances multimodal ultrasound breast lesion segmentation accuracy.
• The proposed framework and dataset offer advancements for clinical diagnosis and treatment planning.
• Improved segmentation accuracy facilitates precise interventions and potentially better patient outcomes.