Zhang Lun, Zhang Junhua
School of Information Science and Engineering, Yunnan University, Kunming, Yunnan, China.
Yunnan Vocational Institute of Energy Technology, Qujing, Yunnan, China.
PeerJ Comput Sci. 2022 Feb 16;8:e873. doi: 10.7717/peerj-cs.873. eCollection 2022.
Ultrasound imaging has been recognized as a powerful tool in clinical diagnosis. Nonetheless, the presence of speckle noise degrades the signal-to-noise of ultrasound images. Various denoising algorithms cannot fully reduce speckle noise and retain image features well for ultrasound imaging. The application of deep learning in ultrasound image denoising has attracted more and more attention in recent years.
In the article, we propose a generative adversarial network with residual dense connectivity and weighted joint loss (GAN-RW) to avoid the limitations of traditional image denoising algorithms and surpass the most advanced performance of ultrasound image denoising. The denoising network is based on U-Net architecture which includes four encoder and four decoder modules. Each of the encoder and decoder modules is replaced with residual dense connectivity and BN to remove speckle noise. The discriminator network applies a series of convolutional layers to identify differences between the translated images and the desired modality. In the training processes, we introduce a joint loss function consisting of a weighted sum of the L1 loss function, binary cross-entropy with a logit loss function and perceptual loss function.
We split the experiments into two parts. First, experiments were performed on Berkeley segmentation (BSD68) datasets corrupted by a simulated speckle. Compared with the eight existing denoising algorithms, the GAN-RW achieved the most advanced despeckling performance in terms of the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and subjective visual effect. When the noise level was 15, the average value of the GAN-RW increased by approximately 3.58% and 1.23% for PSNR and SSIM, respectively. When the noise level was 25, the average value of the GAN-RW increased by approximately 3.08% and 1.84% for PSNR and SSIM, respectively. When the noise level was 50, the average value of the GAN-RW increased by approximately 1.32% and 1.98% for PSNR and SSIM, respectively. Secondly, experiments were performed on the ultrasound images of lymph nodes, the foetal head, and the brachial plexus. The proposed method shows higher subjective visual effect when verifying on the ultrasound images. In the end, through statistical analysis, the GAN-RW achieved the highest mean rank in the Friedman test.
超声成像已被公认为临床诊断中的一种强大工具。尽管如此，散斑噪声的存在会降低超声图像的信噪比。各种去噪算法无法完全减少散斑噪声并很好地保留超声成像的图像特征。近年来，深度学习在超声图像去噪中的应用受到了越来越多的关注。
在本文中，我们提出了一种具有残差密集连接和加权联合损失的生成对抗网络（GAN-RW），以避免传统图像去噪算法的局限性，并超越超声图像去噪的最先进性能。去噪网络基于U-Net架构，包括四个编码器和四个解码器模块。每个编码器和解码器模块都用残差密集连接和BN替换，以去除散斑噪声。判别器网络应用一系列卷积层来识别翻译后的图像与期望模态之间的差异。在训练过程中，我们引入了一个联合损失函数，它由L1损失函数、带逻辑损失函数的二元交叉熵和感知损失函数的加权和组成。
我们将实验分为两部分。首先，在由模拟散斑损坏的伯克利分割（BSD68）数据集上进行实验。与现有的八种去噪算法相比，GAN-RW在峰值信噪比（PSNR）、结构相似性（SSIM）和主观视觉效果方面实现了最先进的去斑性能。当噪声水平为15时，GAN-RW的PSNR和SSIM平均值分别提高了约3.58%和1.23%。当噪声水平为25时，GAN-RW的PSNR和SSIM平均值分别提高了约3.08%和1.84%。当噪声水平为50时，GAN-RW的PSNR和SSIM平均值分别提高了约1.32%和1.98%。其次，在淋巴结、胎儿头部和臂丛神经的超声图像上进行实验。所提出的方法在超声图像验证时显示出更高的主观视觉效果。最后，通过统计分析，GAN-RW在弗里德曼检验中获得了最高的平均排名。
