DFSNE-Net: Deviant feature sensitive noise estimate network for low-dose CT denoising

Highlights

• We assume that the noise is strongly correlated with deviant feature.

• The layers in the U-shaped network are improved to construct DFSNE-Net.

• Balanced loss function and related strategies are proposed to improve GAN training.

• The noise optimization method is proposed to better apply the denoising task.

• Our methods obtains a better Low-dose CT denoising.

Abstract

Background:

Computed tomography (CT) has radiation problems, and high-quality CT scanning is usually accompanied by high doses of radiation that can be harmful to humans, so low-dose CT denoising has received extensive academic attention.

Method:

In this paper, firstly, the concept of deviant features is proposed which lead to the hypothesis of correlation between noise and deviant features. Secondly, in order to estimate the noise in CT, a new method of deviant feature perception and downsampling is proposed. Specifically, the deviant feature perception module based on the multi-scale convolutional cooperative (MSC-DFPM) and Filtering module based on the self-information space attention(SISA-FM) are proposed, and construct the deviant feature sensitive noise estimate network (DFSNE-Net), then a balanced loss function and training strategy adapted to the DFSNE-Net are proposed. Finally, noise distribution normalization based on skewness and kurtosis(SK-NDN) and low credible noise suppression based on confidence interval(CI-LCNS) as the noise optimization methods are proposed to optimize the estimated noise of DFSNE-Net, which is applied to the denoising task of CT.

Conclusions:

Experiments and results demonstrate that our proposed denoising method in this paper obtains a better denoising effect than the current state-of-the-art denoising methods in different evaluation indicator, which proves the hypothesis that noise is strongly correlated with deviant features. It is also proved that the denoising method proposed in this paper can achieve the denoising task for different doses of CT.

Introduction

Computed tomography (CT) is a common examination technique in recently clinical medical field, using X-ray, gamma rays, ultrasound and other related means, in coordination with a highly sensitive detector to perform cross-sectional scans around a part of the body or the whole body, with extremely fast scanning and imaging speed, which can be used for the examination of a variety of diseases. However, the quality of CT can be seriously disturbed and affected by factors such as the CT scanning machine, the CT scanning environment and the patient’s own physical condition.

Although CT provides critical clinical information, radiation poses a potential risk during the scanning process [1], [2]. Over the last decade, radiation dose selection in CT has shown a decreasing trend, for example, typical dose levels for coronary CT angiography have decreased from approximately 12mSv in 2009 [3] to 1.5mSv in 2014 [4], but the reduction in radiation dose in CT examinations inevitably leads to local deviations in HU values [5] with noise and artifacts. Therefore, in order to reduce the damage to the patient’s body and to make more accurate judgments about the patient’s physical condition, improve the image quality, display the image feature more clearly, and make the CT have a better view, the denoising of low-dose CT has gradually become an important issue in the field of CT and has gradually received more academic attention.

For low-dose CT denoising means are divided into the following three main types: Image filtering, iterative reconstruction and image post-processing(both traditional imaging process methods and deep learning methods).

(1) In image filtering algorithms, sinogram processing is usually performed on CT data before image reconstruction. Manduca et al. [6] proposed bilateral filtering combined with a CT noise model, which achieves a better noise resolution tradeoff compared to conventional reconstruction algorithms. Some other typical methods likewise exist including structural adaptive filtering [7] and penalized weighted least squares algorithms [8]. Li et al. [9] proposed a statistical model for sinogram data and developed a penalized likelihood method to suppress quantum noise. Image filtering algorithms are often limited in practical applications due to the difficulty in obtaining projection data.

(2) In iterative reconstruction algorithms, incremental estimation of prior knowledge from the image domain is usually used for denoising. For example, the total variation(TV) [10], [11], [12], [13] incorporates prior knowledge about data noise and image content into the objective function to optimally reconstruct the laminar image, which greatly improves the denoising performance. Similar nonlocal mean methods [14], [15], [16], dictionary learning methods [17], [18] and some similar other techniques [19], [20], [21] have also been formulated. However, a large amount of computational resources are used on the iterative projection/inverse projection step, which also requires a long processing time on specialized hardware and the availability of CT projection data.

(3) In image post-processing algorithm, the reconstructed CT are usually post-processed directly, instead of directly processing the original projection data. Initially, researchers tried classical image processing algorithms, similar to nonlocal mean filtering [22], [23], [24], dictionary learning-based methods [25], [26], block matching algorithms [27], [28], [29], and diffusion filters [30], which are much more computationally efficient than iterative reconstruction algorithms. However, the noise with in the reconstructed images after low-dose CT is often unevenly distributed, and processing by these algorithms does not solve the problem of low-dose CT denoising, so it is not widely used.

(4) In image post-processing algorithm, the use of deep neural networks for low-dose CT denoising has gradually become an important approach in recent years. Convolutional Neural Network (CNN) is a class of deep neural networks with nonlinear and convolutional filters that have proven to be very effective in a variety of tasks such as image classification, denoising [31], [32], [33], and super-resolution [34]. Early research in denoising focused on CNN structure optimization and adaption, such as RED-CNN [35] and wavelet networks [36]. However, the use of Root Mean Square Error Loss in these methods can produce excessively smooth images for denoised images, Johnson et al. proposed a perceptual loss based on a pre-trained VGG model [37], [38] to solve this problem. Generative adversarial network (GAN) [39], [40] was subsequently introduced to implement low-dose CT denoising, which transformed the image denoising problem into a high-quality CT generation problem, reducing the pixel-level regression loss [41], [42], [43], [44] and improving the quality of the denoised images.

The rest of this paper is organized as follows. Part 2 realizes the deviant analysis of noise through mathematical model. Part 3 introduces the key methods used in our proposed denoising model. Part 4 reports the experiment details and results. Part 5 discusses some related issues, gives the future development direction and make a conclusion.