BCT-Net: semantic-guided breast cancer segmentation on BUS.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate breast tumor segmentation is crucial for cancer diagnosis and treatment.
• Ultrasound imaging is widely used but presents segmentation challenges like low contrast and blurred boundaries.
• Existing segmentation methods struggle with the complexities of breast ultrasound images.

Purpose of the Study:

• To develop an advanced deep learning network for accurate breast tumor segmentation in ultrasound images.
• To improve segmentation performance by integrating Convolutional Neural Network (CNN) and transformer architectures.
• To enhance semantic information and feature capture for more precise tumor delineation.

Main Methods:

• Proposing BCT-Net, a hybrid CNN-transformer network with a dual-level attention mechanism.
• Redefining the skip connection module to better integrate features across network levels.
• Utilizing a classification task as an auxiliary task with supervised contrastive learning.
• Employing a hybrid objective loss function combining cross-entropy and contrastive learning losses.

Main Results:

• BCT-Net achieved high segmentation precision with Pre (86.12%) and DSC (88.70%) indices.
• The network demonstrated high accuracy on the BUSI dataset for breast ultrasound images.
• The proposed methods effectively addressed challenges like low contrast and blurred boundaries.

Conclusions:

• BCT-Net offers a robust and accurate solution for breast tumor segmentation in ultrasound imaging.
• The integration of CNN, transformers, and attention mechanisms significantly enhances segmentation performance.
• This approach holds promise for improving clinical diagnosis and treatment planning for breast cancer.