[论文解读] CT Image Denoising with Perceptive Deep Neural Networks

Qingsong Yang, Pingkun Yan+2 more|arXiv (Cornell University)|Feb 22, 2017

Medical Imaging Techniques and Applications参考文献 6被引用 57

一句话总结

本文提出基于感知损失的低剂量CT去噪，使用由预训练VGG网络特征引导的CNN，在结构保留方面优于像素级损失。

ABSTRACT

Increasing use of CT in modern medical practice has raised concerns over associated radiation dose. Reduction of radiation dose associated with CT can increase noise and artifacts, which can adversely affect diagnostic confidence. Denoising of low-dose CT images on the other hand can help improve diagnostic confidence, which however is a challenging problem due to its ill-posed nature, since one noisy image patch may correspond to many different output patches. In the past decade, machine learning based approaches have made quite impressive progress in this direction. However, most of those methods, including the recently popularized deep learning techniques, aim for minimizing mean-squared-error (MSE) between a denoised CT image and the ground truth, which results in losing important structural details due to over-smoothing, although the PSNR based performance measure looks great. In this work, we introduce a new perceptual similarity measure as the objective function for a deep convolutional neural network to facilitate CT image denoising. Instead of directly computing MSE for pixel-to-pixel intensity loss, we compare the perceptual features of a denoised output against those of the ground truth in a feature space. Therefore, our proposed method is capable of not only reducing the image noise levels, but also keeping the critical structural information at the same time. Promising results have been obtained in our experiments with a large number of CT images.

研究动机与目标

• 通过去噪低剂量获取来推动CT中的辐射剂量下降而不过度平滑。
• 提出一种使用感知损失而非像素级MSE的CNN去噪框架。
• 证明感知损失能够保留结构细节并提升定性图像质量。
• 在传统方法和标准CNN去噪方法上进行对比，显示感知损失的优势。

提出的方法

• 使用一个8层、3x3卷积核的CNN来学习去噪。
• 定义基于预训练VGG网络的特征图的感知重建损失。
• 将去噪输出与真值图像的VGG特征图之间的MSE作为损失（phi_i项）。
• 尝试VGG层relu1_1、relu3_1和relu3_4作为感知辅助（CNN-VGG11、CNN-VGG31、CNN-VGG34）。
• 从MSE训练的CNN（CNN-MSE）初始化CNN-VGG网络并进行微调。
• 在具有不同噪声指标和重建方法（FBP vs MBIR/VEO）的人体尸体CT数据上进行训练。
• 训练结束后，每张图像去噪耗时小于5秒。

实验结果

研究问题

• RQ1 感知特征为基础的损失是否相对于像素级损失可以提升CT去噪质量？
• RQ2 感知损失是否在降低噪声的同时保留诊断性结构细节？
• RQ3 CNN-VGG变体在图像清晰度和与真值的相似性方面有何差异？
• RQ4 感知型CNN在定量指标和视觉质量上是否具备与BM3D及先前CNN方法的竞争力？

主要发现

• 使用感知损失训练的CNN模型（CNN-VGG31、CNN-VGG34）在视觉细节和对比度方面优于CNN-MSE。
• CNN-MSE在PSNR/SSIM上高于部分感知变体，但感知变体在视觉上更具吸引力且更能保留细节。
• CNN-VGG31和CNN-VGG34在相对于ground-truth VEO10NI的病灶勾画方面提供了最佳分辨，同时提升了整体图像外观。
• 定量指标显示CNN-MSE在PSNR/SSIM上优于感知变体，但感知变体在保留图像结构方面表现更好。
• BM3D的表现因图像而异，存在一些平滑和残留伪影。"
• Table 1 报告 PSNR/SSIM: FBP30NI 27.1544/0.8018; CNN-MSE 31.1135/0.9351; CNN-VGG11 31.1239/0.9348; CNN-VGG31 30.6462/0.9260; CNN-VGG34 30.2154/0.9159; BM3D 28.7405/0.9026。