LTM-UNet: Linear Transformer-Mamba with Attention-Based U-Net for Context-Aware Breast Ultrasound Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast lesion segmentation is crucial for diagnosis.
• Existing deep learning models struggle with global context and fine details.

Purpose of the Study:

• To develop a novel segmentation model for improved context-aware dense segmentation of ultrasound images.
• To address limitations in current deep learning approaches for medical image analysis.

Main Methods:

• Proposed LTM-UNet, integrating a vision transformer encoder with a state-space model decoder in a U-Net framework.
• Utilized attention-guided skip-fusion to preserve spatial details and minimize the semantic gap.
• Employed a direction-aware decoder for efficient long-range dependency capture.

Main Results:

• LTM-UNet achieved a Dice score of 82.41% on the BUSI dataset and 86.62% on Dataset B (UDIAT).
• The model outperformed existing segmentation methods in Dice score and Intersection-over-Union (IoU) metrics.
• Demonstrated effectiveness on benchmark ultrasound medical imaging datasets.

Conclusions:

• LTM-UNet effectively captures structural details and contextual information for superior segmentation performance.
• The integration of transformer encoding, attention fusion, and state-space decoding enhances segmentation accuracy.
• The proposed method offers a promising advancement in automated breast lesion analysis.