Breast Ultrasound Image Segmentation Using Multi-branch Skip Connection Search.

Abstract

To reduce the cost of designing neural networks and improve the accuracy of breast ultrasound image segmentation, an encoder-decoder neural network architecture search method is proposed, tailored for constructing segmentation models automatically. Initially, a multi-branch skip connection module is designed in which each branch utilizes distinct operations to extract features of varying scales and types from subsets of channels. Subsequently, a learnable operation weight search strategy is introduced that employs Gumbel-Softmax for reparameterizing discrete operation weights. This strategy explores optimal operations within the multi-branch skip connection module through both shared and non-shared methodologies. The candidate neural networks incorporate encoder-decoder block pairs that utilize the Swin Transformer from Swin-Unet and convolutional blocks from TransUNet, respectively. Experimental results demonstrate that the method identifies the optimal encoder-decoder model in approximately two hours. The automatically constructed model achieves superior segmentation accuracy, with Dice scores of approximately 85.94% and 84.44% on the BUSI and OASBUD datasets, respectively. It outperforms state-of-the-art (SOTA) methods such as AAU-Net, SK-U-Net, and TransUNet. High-precision segmentation results offer clear localization of lesion boundaries, thereby reducing the risk of missed diagnoses. The model's quantitative metrics, such as lesion area and morphology, can be seamlessly incorporated into diagnostic reports, facilitating the development of personalized treatment plans.