A Breast Lesion Segmentation Method Based on Radio Frequency Ultrasound Signals

Accurate breast detection and segmentation methods can improve the effectiveness of detection and diagnosis of breast disease, while simultaneously alleviating the workload of medical practitioners. In recent years, numerous methods have emerged for segmenting breast lesions. However, most of them rely on B-mode ultrasound images and exhibit limited understanding of the primary data. To improve the accuracy of segmentation, a segmentation algorithm based on the original ultrasound RF signal is proposed in this paper. The algorithm first uses the MimickNet technique for noise reduction and compression of the original radio frequency (RF) signal. Then, the boundary prediction is accomplished using the Visual Geometry Group 16 (VGG16) neural network as a boundary probability detector. To mitigate the error introduced by the binarization of the boundary probability matrix, a negative feedback-based optimizer is utilized. In the experiments, medical ultrasound images from the publicly available dataset OASBUD are segmented using the algorithm in this paper. The results are compared with those by the U-net method, threshold method, watershed algorithm and texture-based algorithm. It turns out that the algorithm in this paper has great accuracy and stability in noise reduction, compression processing, boundary prediction and accuracy maintenance.