FDDL-Net: frequency domain decomposition learning for speckle reduction in ultrasound images

Image decomposition is a useful operation that benefits a number of low-level vision tasks. However, this conventional wisdom is not well studied in deep learning, and almost no existing deep learning-based methods consider the fact that the extracted feature map from a convolution layer consists of different frequency information. We propose an end-to-end frequency domain decomposition learning network (FDDL-Net) to remove speckle noise from ultrasound images. FDDL-Net leverages frequency domain decomposition at the feature level to learn structure and detail information from ultrasound images via an interactive dual-branch framework. According to the properties of speckle noise, the median filter is utilized in the high-frequency branch of the network to remove the noise effectively. In addition, information from the two branches is exchanged interactively, so that valuable features from different frequencies are fully exploited for speckle reduction. Compared with state-of-the-art methods, FDDL-net demonstrates superior noise reduction and feature preservation (0.89 and 30.92 for SSIM and PSNR metrics respectively), attributing to the dual-branch interaction of the network.