Abstract
Most of the existing medical image denoising methods focus on estimating either the image or the residual noise. Moreover, they are usually designed for one specific noise with a strong assumption of the noise distribution. However, not only the random independent Gaussian or speckle noise but also the structurally correlated ring or stripe noise is ubiquitous in various medical imaging instruments. Explicitly modeling the distributions of these complex noises in the medical image is extremely hard. They cannot be accurately held by the Gaussian or mixture of the Gaussian model. To overcome the two drawbacks, in this paper, we propose to treat the image and noise components equally and convert the image denoising task into an image decomposition problem naturally. More precisely, we present a two-stage deep convolutional neural network (CNN) to model both the noise and the medical image simultaneously. On the one hand, we utilize both the image and noise to separate them better. On the other hand, the noise subnetwork serves as a noise estimator which guides the image subnetwork with sufficient information about the noise, thus we could easily handle different noise distributions and noise levels. To better cope with the gradient vanishing problem in this very deep network, we introduce both the short-term and long-term connections in the network which could promote the information propagation between different layers efficiently. Extensive experiments have been performed on several kinds of medical noise images, such as the computed tomography and ultrasound images, and the proposed method has consistently outperformed state-of-the-art denoising methods.
Author Keywords
- Convolutional neural network (CNN),
- image decomposition,
- image denoising,
- medical image
IEEE Keywords
- Mathematical model,
- Computed tomography,
- Image decomposition,
- Image denoising,
- Medical diagnostic imaging,
- Noise reduction
Introduction
The medical noises would obviously increase the uncertainties in the measurement procedures and degrade the quality of the images seriously, which makes them diagnostically unusable. Numerous image denoising methods have been proposed in the past decades. In this paper, from a more general perspective, we define the noise (more precisely, we could name “it” here as the artifacts) as anything that is not expected to be presented in the medical images. The noises have different appearances in different imaging instruments and can be broadly classified into two categories: the independent random noise and the structurally correlated/fixed pattern noise. In this paper, for the random noise, we mainly focus on the additive Gaussian random noise in computed tomography (CT) image and the multiplicative speckle noise in the ultrasound image. For the structural noise, we mainly focus on the line pattern stripe noise in the scanning electron microscope (SEM) image and the circle pattern ring noise in the CT image, as shown in Fig. 1.
The goal of this work is to suppress all these artifacts or the mixture of them via a unified image processing method. However, there are three main limitations of the optimization-based methods. First of all, the optimization methods need explicitly model. For the real-world medical images, where the noises are much more complex and vary with different appearances such as the structural or mixed noise, it is very hard to figure out the corresponding mathematical formulations precisely.
The approximations via a mixture of Gaussian (MoG) or Laplacian also fail to model the structural noise. Second, most of the previous optimization-based methods mainly utilize a predefined prior. Such a hand-crafted prior is definitely not suitable for multi-modality medical images. Last but not least, the computational load is heavy due to the iteration and complex operation, which limits their potential for real-time application. Recently, CNN has been widely used for low-level image tasks, such as image denoising [19], deblurring [20], and super-resolution [21].
Moreover, the learned prior to the training pairs make the network more adaptive for specific images. At last, due to the simple operation of the network, its forward process is extremely fast which makes it quite suitable for real-time applications.
Conclusion
In this paper, we propose to remove the noises in medical images from the image decomposition perspective. Different from previous works, the noise component and image component are treated equally in our work, and a two-stage convolutional neural network is proposed to model both the image and noise simultaneously. Instead of explicitly modeling the complex distribution of various noises and multi-modality medical images, the deep model could automatically figure out the distribution of the specific noise and image from a data-driven viewpoint. The proposed cascaded CNN model benefits us to handle different noise categories and noise levels adaptively. To facilitate the training, we introduce both the short-term and long-term connections in the network for better information propagation. Moreover, we apply the finetune strategy to alleviate the lack of medical images issue. Extensive simulated and real medical image datasets have been tested. Experimental results demonstrate that the proposed method is very effective for various noises, and outperforms the state-of-the-art methods. Our work shows that CNN is a powerful tool for modeling the noises and multi-modality images with fast test speed. We believe other low-level image processing problems such as the deblurring and super-resolution tasks could also benefit from the deep model. Moreover, it is interesting to see more advanced deep models for medical image analysis. For example, 3D CNN could well handle multi-slice data with temporal information.
About KSRA
The Kavian Scientific Research Association (KSRA) is a non-profit research organization to provide research / educational services in December 2013. The members of the community had formed a virtual group on the Viber social network. The core of the Kavian Scientific Association was formed with these members as founders. These individuals, led by Professor Siavosh Kaviani, decided to launch a scientific / research association with an emphasis on education.
KSRA research association, as a non-profit research firm, is committed to providing research services in the field of knowledge. The main beneficiaries of this association are public or private knowledge-based companies, students, researchers, researchers, professors, universities, and industrial and semi-industrial centers around the world.
Our main services Based on Education for all Spectrum people in the world. We want to make an integration between researches and educations. We believe education is the main right of Human beings. So our services should be concentrated on inclusive education.
The KSRA team partners with local under-served communities around the world to improve the access to and quality of knowledge based on education, amplify and augment learning programs where they exist, and create new opportunities for e-learning where traditional education systems are lacking or non-existent.
FULL Paper PDF file:
Two-Stage Convolutional Neural Network for Medical Noise Removal via Image DecompositionBibliography
author
Year
2020
Title
Two-Stage Convolutional Neural Network for Medical Noise Removal via Image Decomposition
Publish in
in IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 6, pp. 2707-2721, June 2020
Doi
10.1109/TIM.2019.2925881
PDF reference and original file: Click here
Somayeh Nosrati was born in 1982 in Tehran. She holds a Master's degree in artificial intelligence from Khatam University of Tehran.
