摘要：

基于卷积神经网络(CNN)的图像去噪方法能有效去除低剂量计算机断层扫描(CT)图像伴随的伪影和噪声, 从而确保CT设备输出高质量图像同时降低辐射, 这对患者健康和医学诊断具有重要意义。为了进一步提高低剂量CT图像的质量, 提出一种小波域去噪网络MDTNet。首先, 基于双树复小波变换(DTCWT)构造多级编解码去噪网络, 在多个尺度上提取特征以保留更多高频细节; 然后, 利用扩展的像素重排技术替代卷积上下采样, 实现多级输入和特征融合, 从而降低计算复杂度; 最后, 通过大量训练找到最佳的去噪模型, 即二级MDTNet配合LeGall滤波器和Qshift_b滤波器, 并选择较大尺寸的CT图像作为训练数据。使用AAPM数据集评估MDTNet的性能, 实验结果表明, MDTNet能有效去除条纹状伪影和噪声, 在定量和定性评估中性能均优于同类型去噪方法。与FWDNet相比, 对于1 mm的切片, MDTNet的平均峰值信噪比(PSNR)和结构相似性指数(SSIM)分别提高了0.088 7 dB和0.002 4;对于3 mm的切片, 分别提升了0.144 3 dB和0.003。对于单张512×512像素的低剂量CT图像去噪, MDTNet在GPU上仅需0.193 s。MDTNet在保持高效率的同时保留了更多的高频细节, 能够为低剂量CT图像去噪提供一种新的框架。

关键词: 低剂量CT图像, 图像去噪, 卷积神经网络, 双树复小波变换, 像素重排

Abstract:

Based on the Convolutional Neural Network (CNN), image denoising methods can effectively remove the artifacts and noise associated with low-dose Computed Tomography (CT), thereby ensuring high-quality output while minimizing radiation exposure. This information is of great significance for patient health and medical diagnosis. This study proposes a novel denoising network called MDTNet to enhance the quality of low-dose CT images. In this approach, multilevel encoder-decoder denoising networks are constructed using a Dual-Tree Complex Wavelet Transform (DTCWT), which enables the preservation of high-frequency details. Moreover, a pixel shuffle was employed to facilitate multi-level input and feature fusion, resulting in significantly reduced computation and memory complexity. In addition, by training a set of residual mappings in the wavelet domain, optimal denoising performance was achieved using a two-level MDTNet with LeGall and Qshift_b filters. The effectiveness of the MDTNet was evaluated on the 2016 NIH-AAPM-Mayo Clinic low-dose CT grand challenge dataset. The experimental results demonstrate that MDTNet outperforms state-of-the-art denoising methods in quantitative and qualitative evaluations. Specifically, compared with the FWDNet, on 1 mm slices, MDTNet improved the average Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) by 0.088 7 dB and 0.002 4, respectively; on a 3 mm slice, the increase was 0.144 3 dB and 0.003, respectively. Moreover, MDTNet processed 512×512 low-dose CT images on a Graphical Processing Unit (GPU) in 0.193 s. Preserving high-frequency details while maintaining efficiency, MDTNet presents an innovative framework for denoising low-dose CT images.

Key words: low-dose CT image, image denoising, Convolutional Neural Network(CNN), Dual-Tree Complex Wavelet Transform(DTCWT), pixel shuffle