RFImageNet framework for segmentation of ultrasound images with spectra-augmented radiofrequency signals

Abstract

Computer-aided segmentation of medical ultrasound images assists in medical diagnosis, promoting accuracy and reducing the burden of sonographers. However, the existing ultrasonic intelligent segmentation models are mainly based on B-mode grayscale images, which lack sufficient clarity and contrast compared to natural images. Previous research has indicated that ultrasound radiofrequency (RF) signals contain rich spectral information that could be beneficial for tissue recognition but is lost in grayscale images. In this paper, we introduce an image segmentation framework, RFImageNet, that leverages spectral and amplitude information from RF signals to segment ultrasound image. Firstly, the positive and negative values in the RF signal are separated into the red and green channels respectively in the proposed RF image, ensuring the preservation of frequency information. Secondly, we developed a deep learning model, RFNet, tailored to the specific input image size requirements. Thirdly, RFNet was trained using RF images with spectral data augmentation and tested against other models. The proposed method achieved a mean intersection over union (mIoU) of 54.99% and a dice score of 63.89% in the segmentation of rat abdominal tissues, as well as a mIoU of 63.28% and a dice score of 68.92% in distinguishing between benign and malignant breast tumors. These results highlight the potential of combining RF signals with deep learning algorithms for enhanced diagnostic capabilities.