RFF-Net: A Refined Feature Feedback Network for Muscle Ultrasound Image Segmentation with Feature Subtraction and Deep Supervision

Ultrasound imaging is being used as a new diagnostic tool for identifying sarcopenia. However, the low contrast characteristics of ultrasound images and significant scale variations in muscle areas pose certain challenges to segmentation. Therefore, we propose a segmentation network called RFF-Net to automatically and accurately segment the muscle region in ultrasound images. RFF-Net comprises three novel components: (1) A multi-scale feature subtraction module (MFS) is designed to weaken redundant features to achieve accurate segmentation; (2) A refinement feature feedback module (RFF) is proposed to extract ambiguous boundary features to improve segmentation integrity; (3) A multi-resolution deep supervision module (MDS) is introduced to perform feature selection for different resolution features generating from decoder to improve segmentation accuracy. Experiments on both private and public datasets show our method achieves much higher segmentation accuracy than related methods.