Self-supervised ultrasound image denoising based on weighted joint loss

Speckle noise is an important degradation factor of ultrasound imaging, which affects its clinical application. Self-supervised denoising methods based on deep learning have been developing rapidly. However, most of them primarily address spatially independent noise and are not suitable for removing spatially correlated noise. In addition, as a difficult problem in the image denoising task, balancing noise removal and preserving image details has also been the research focus of various denoising methodologies. To address the above problems, this paper proposes a self-supervised ultrasound image denoising algorithm that utilizes a sampling method to construct sub-image pairs as supervision and uses different denoisers for joint training with a novel weighted joint loss. For the input raw noisy image, it is first chunked, then pixel points on the diagonal of the image chunks are randomly sampled and formed into subsampled image pairs as supervision to train the network. Considering the presence of regions in the image with different texture complexity, a joint model based on blind-neighborhood network and U-Net is used as denoising network in the training stage, which strives to remove the noise while preserving the image details. Additionally, this paper uses the standard deviation of local image blocks as the measure of texture complexity and transforms them to adaptive coefficients. In the training process, we use adaptive coefficients to construct the weighted joint loss functions for adjusting the degree of influence of two denoisers on model. In comparison with the self-supervised denoising algorithm Neighbor2Neighbor, the supervised denoising methods RNAN and Restormer, and non-learning denoising methods BM3D and OBNLM, the proposed method achieves better denoising effects on both synthetic images and real ultrasound images.