Ultrasound Radiofrequency Image Improves the Tissue Segmentation Performance of Deep Learning Models

The past few years have witnessed remarkable advancements in deep learning-based algorithms for ultrasound image segmentation. However, traditional B-mode ultrasound image only captures the amplitude envelope information of the ultrasound signal, leading to limited image resolution and contrast. To enhance the understanding of ultrasound images by deep learning algorithms, current efforts primarily focus on incorporating additional ultrasound image information. For instance, techniques such as elastic ultrasound imaging, contrast-enhanced imaging, and quantitative ultrasound parameters effectively improved the performance of deep learning models. Nevertheless, most of these methods require extracting supplementary feature parameters and constructing a new ultrasound parameter map for input into neural networks, which increases the complexity of models and consumes substantial computational resources. In this study, we propose a novel ultrasound imaging mode called ultrasound radiofrequency (RF) image, which preserves the time-frequency information from the original RF signal to enhance deep learning-based segmentation tasks for ultrasound images. Experimental results demonstrate that RF images exhibit significant improvements over B-mode ultrasound images in terms of deep learning segmentation. The performance of multiple deep-learning image segmentation models is improved without designing additional model structures.