基于卷积神经网络的超声图像散斑去噪算法

为了提高超声图像质量，解决传统去噪算法在抑制散斑噪声和保留超声图像纹理特征方面的难题，提出一种基于卷积神经网络的超声图像散斑去噪算法DSCNN（De-speckling CNN）。本文提出的算法利用卷积神经网络强大的拟合能力来学习从超声图像到其相应的高质量图像的复杂映射，同时，通过改进损失函数的方式来减少去噪过程中纹理信息的损失和细节的模糊。不同于以往简单地假设超声散斑噪声为乘性噪声，本文利用基于超声图像采集模型和散斑噪声形成模型的模拟超声成像技术为去噪模型生成更贴合真实超声图像的训练数据，解决深度学习方法训练数据匮乏以及在临床上无法获得与超声图像空间配准作为标签的无噪声图像的难题。通过与其他具有代表性的超声图像去噪算法比较，经DSCNN去噪后的超声图像无论在视觉效果还是图像质量评价指标上都取得了更好的结果，其中SSIM达到0.856 9，在文中所有方法中最高。

Convolutional neural network-based de-speckling algorithm for ultrasound images

Abstract: In order to improve the quality of ultrasound images and solve the problems of how to maintain the balance between suppressing speckle noise and retaining the texture features in ultrasound images, a de-speckling algorithm based on convolutional neural network (DSCNN) is proposed. The proposed algorithm utilizes the powerful fitting ability of CNN to learn the complex mapping from the ultrasound image to its corresponding high-quality image. Meanwhile, it reduces the loss of texture information and the blurred details during denoising by advanced loss function. Different from previous assumption that the speckle noise is a kind of multiplicative noise, the simulated ultrasound imaging technology based on the ultrasound image acquisition model and the speckle noise formation model is used to generate more realistic ultrasound images which are taken as the training data for denoising model, thus solving the problems of lack of training data and overcoming difficulties in obtaining a noise-free image which is used as a label in spatial registration with the ultrasound image. Compared with the images denoised by other typical denoising algorithms, the ultrasound image denoised by DSCNN wins higher scores both in visual effect and image quality assessment, and SSIM even reaches 0.856 9, which is highest among all methods discussed in the study.