Breast Ultrasound Image Segmentation Using Multi-branch Skip Connection Search

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast ultrasound image segmentation is crucial for early cancer detection and treatment planning.
• Current neural network design for segmentation is costly and time-consuming.
• Improving segmentation accuracy can reduce missed diagnoses and aid personalized medicine.

Purpose of the Study:

• To develop an automated neural network architecture search method for breast ultrasound image segmentation.
• To reduce the cost and improve the accuracy of segmentation model design.
• To enhance the localization of lesion boundaries for clinical decision support.

Main Methods:

• Proposed a novel encoder-decoder neural network architecture search method.
• Introduced a multi-branch skip connection module with distinct feature extraction operations.
• Employed a learnable operation weight search strategy using Gumbel-Softmax for optimization.
• Integrated Swin Transformer and convolutional blocks within candidate models.

Main Results:

• Identified the optimal encoder-decoder model in approximately two hours.
• Achieved superior segmentation accuracy with Dice scores of ~85.94% (BUSI) and ~84.44% (OASBUD).
• Outperformed state-of-the-art methods including AAU-Net, SK-U-Net, and TransUNet.

Conclusions:

• The automated method efficiently designs high-performance segmentation models.
• The superior accuracy enhances lesion boundary localization, reducing diagnostic errors.
• Quantitative metrics from segmentation can support personalized treatment planning.