A novel ultrasound image denoising algorithm based on Neighbor2Neighbor unsupervised learning model

Abstract Ultrasound images are widely adopted in medical diagnosis, thus whose the quality and accuracy are crucial. The mainstream supervised deep learning denoisation methods generally require noise-clean image pairs of samples, which is not feasible in real ultrasound scenarios. Due to these limitations of independent noise, current unsupervised denoizing methods like Neighbor2Neighbor cannot address correlated noise in ultrasound images. Meanwhile, due to the random neighborhood downsampler, this method results in pixel loss. This work proposes an effective Neighbor2Neighbor algorithm to deal with the pixel loss problem by reconstructing the image through a refined downsampling strategy. Moreover, the structural similarity function and the regularization term are introduced into the loss function, which can suppress the independent and correlated noises more effectively. After experimental comparison with the state-of-the-art baseline algorithms including BM3D, Noise2Noise, Noise2Void, Neighbor2Neighbor, the proposed algorithm has excellent performance in PSNR, MSSIM, PSLM and EPI.