【摘要】 为进一步提升医学图像的视觉质量，针对DnCNN算法的局限性和医学图像的特征，采用改进的深度卷积神经网络算法进行图像去噪。改进方法应用多尺寸卷积核提取医学图像不同尺度特征，增加深度网络对医学图像的适应性；改进常规网络中激活函数和损失函数的定义方式，从而更好地保护输出结果中的边缘/纹理细节信息；增加一个跳跃连接，提升网络的训练速度和算法的收敛精度。仿真模拟结果表明，相比DnCNN算法、ID-CNN算法、BM3D算法和曲波变换去噪算法，该改进网络具有更好图像细节保持以及更好的去噪效果，图像的峰值信噪比平均提升56%以上，结构相似度平均提升至0.881。改进的深度卷积网络去噪效果好、效率高，在医学图像处理中具有较强的推广性。更多还原

【Abstract】 In order to further improve the visual quality of medical images, aimed at the limitations of de-noising convolutional neural networks(DnCNN) algorithm and the characteristics of medical image, an improved deep convolution neural network algorithm is used for image de-noising. The improved method used a multi-core convolution layer to extract different scale features of medical image, which increased the adaptability of depth network to medical image. The definition of activation function and loss function in conventional network was improved, so as to better protect the edge/texture details in the output results. A jump link was added to improve the training speed of network and the convergence accuracy of algorithm. The simulation results show that compared with the DnCNN, ID-CNN(Iteration dilated convolution neural networks), BM3 D(Block-Matching and 3 D filtering), curvelet transform filtering, the improved method has better image detail preservation, and better de-noising effect, the peak signal-to-noise ratio of the image is improved by more than 56% on average, structural similarity increased to 0.881 on average. The improved method has a good effect of medical image de-noising and high efficiency, which can be easily popularized in medical image processing.更多还原