Cross-Image Dependency Modeling for Breast Ultrasound Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Breast ultrasound lesion segmentation is challenging due to lesion variability, unclear boundaries, and image artifacts.
• Existing methods often overlook cross-image dependencies crucial for improving segmentation accuracy with limited data and noise.

Purpose of the Study:

• To introduce BUSSeg, a novel deep network for automated breast ultrasound lesion segmentation.
• To enhance segmentation by incorporating both within- and cross-image long-range dependency modeling.

Main Methods:

• Developed a Cross-Image Dependency Module (CDM) using complete spatial features and a Cross-Image Dependency Loss (CDL) for consistent feature expression and noise reduction.
• Implemented a Parallel Bi-Encoder Architecture (PBA) combining Transformer and CNN to capture within-image dependencies.
• BUSSeg integrates CDM and PBA for comprehensive lesion segmentation.

Main Results:

• The proposed CDM effectively captures semantic dependencies between images, mitigating noise and improving feature representation.
• The PBA enhances the model's ability to capture within-image long-range dependencies, providing richer features for the CDM.
• BUSSeg demonstrated superior performance over state-of-the-art methods on two public breast ultrasound datasets.

Conclusions:

• BUSSeg offers a significant advancement in automated breast ultrasound lesion segmentation.
• The integration of within- and cross-image dependency modeling is effective for handling challenges in ultrasound imaging.
• The proposed methods provide a robust framework for improving diagnostic accuracy in breast cancer screening.