Multimodal U-Net Breast Cancer Tumor Algorithm Based on Radio Frequency (RF) Data and Ultrasound Images
Zeping Chen, Bingbing He, Zhouyuan Liang, Hong X. Tang, Yangchen Fu, Yufeng Zhang · 2023
Breast cancer is a common cancer among the women worldwide,early screening is crucial for timely treatment and improved therapeutic effect. The unimodal algorithm has difficulty in accurately segmenting tumour regions when processing ultrasound images with poor imaging quality. Therefore, a multimodal M-UNet breast tumor segmentation algorithm based on radio frequency (RF) data and ultrasound images is proposed. Breast tumor ultrasound images and RF data are used as multimodal inputs, and the segmentation of breast tumor ultrasound images is performed by combining Multi-Head Attention in Transformer with U-Net network. The proposed algorithm is tested on the Open Access Series of Breast Ultrasonic Data (OASBUD), a publicly available tumor dataset from the Ultrasound Department of the Institute of Basic Technologies of the Polish Academy of Sciences, and yields a Dice coefficient of 66.0% and a cross-merge ratio coefficient (IoU) of 52.7%. Compared with the unimodal U-Net method, the proposed algorithm improves the Dice coefficient and IoU by 8% and 3% respectively. Therefore, this method is expected to play an important auxiliary role in the early screening of breast cancer.