"Low-Dose CT" denoising paper

foreword

1.<Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network (RED-CNN)>
2.<Wavelet Domain Residual Network (WavResNet) for Low-Dose X-ray CT Reconstruction>
3.<Structure-sensitive Multi-scale Deep Neural Network for Low-Dose CT Denoising>
4.<Deep Convolutional Framelet Denosing for Low-Dose CT via Wavelet Residual Network>
In fact, it was because I wanted to do some work on medical images at first, but later I felt that it was not easy to do, because the performance was already better, and it was easier than natural image denoising, and then to change the network model structure, very few Room for improvement.

1. "Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network (RED-CNN)"

1. Abstract

Abstract—Considering the potential danger of X-rays to patients, low-dose CT has attracted great interest in the field of medical imaging. Currently, mainstream low-dose CT methods include vendor-specific sinusoidal domain filtering and iterative reconstruction algorithms, but they require access to raw data in a format that is opaque to most users. Due to the difficulty of modeling statistical features in the image domain, existing methods that directly process reconstructed images cannot remove image noise well while preserving structural details. Inspired by the idea of ​​deep learning, here we concatenate autoencoders, deconvolutional networks and shortcuts into a residual encoder-decoder convolutional neural network (RED-CNN) for low-dose CT imaging. After patch-based training, the proposed RED-CNN has competitive advantages over state-of-the-art methods in both simulation and clinical cases. In particular, our method is well evaluated in terms of noise suppression, structure preservation and lesion detection.

2. Network Model



code show as below: Link

import os
import numpy as np
import torch.nn as nn

class RED_CNN(nn.Module):
    def __init__(self, out_ch=96):
        super(RED_CNN, self).__init__()
        self.conv1 = nn.Conv2D(1, out_ch, kernel_size=5, stride=1, padding=0)
        self.conv2 = nn.Conv2D(out_ch, out_ch, kernel_size=5, stride=1, padding=0)
        self.conv3 = nn.Conv2D(out_ch, out_ch, kernel_size=5, stride=1, padding=0)
        self.conv4 = nn.Conv2D(out_ch, out_ch, kernel_size=5, stride=1, padding=0)
        self.conv5 = nn.Conv2D(out_ch, out_ch, kernel_size=5, stride=1, padding=0)

        self.tconv1 = nn.ConvTranspose2d(out_ch, out_ch, kernel_size=5, stride=1, padding=0)
        self.tconv2 = nn.ConvTranspose2d(out_ch, out_ch, kernel_size=5, stride=1, padding=0)
        self.tconv3 = nn.ConvTranspose2d(out_ch, out_ch, kernel_size=5, stride=1, padding=0)
        self.tconv4 = nn.ConvTranspose2d(out_ch, out_ch, kernel_size=5, stride=1, padding=0)
        self.tconv5 = nn.ConvTranspose2d(out_ch, 1, kernel_size=5, stride=1, padding=0)

        self.relu = nn.ReLU()

    def forward(self, x):
        # encoder
        residual_1 = x
        out = self.relu(self.conv1(x))
        out = self.relu(self.conv2(out))
        residual_2 = out
        out = self.relu(self.conv3(out))
        out = self.relu(self.conv4(out))
        residual_3 = out
        out = self.relu(self.conv5(out))
        # decoder
        out = self.tconv1(out)
        out += residual_3
        out = self.tconv2(self.relu(out))
        out = self.tconv3(self.relu(out))
        out += residual_2
        out = self.tconv4(self.relu(out))
        out = self.tconv5(self.relu(out))
        out += residual_1
        out = self.relu(out)
        return out

2. Wavelet Domain Residual Network (WavResNet) for Low-Dose X-ray CT Reconstruction

1. Summary

Abstract—Model-based iterative reconstruction (MBIR) algorithms for low-dose X-ray CT are computationally complex due to the repeated use of forward and reverse projections. Inspired by the success of deep learning in computer vision applications, we recently proposed a deep convolutional neural network (CNN) for low-dose X-ray CT and won the second place in the 2016 AAPM Low-dose CT Grand Prix. However, some textures were not fully restored, which is unfamiliar to radiologists. To address this problem, we propose a direct residual learning method on the directional wavelet domain to solve the problem and improve its performance. In particular, the new network estimates the noise of each input wavelet transform and then obtains denoised wavelet coefficients by subtracting the noise from the input wavelet transform band. Experimental results confirm that the proposed network has significantly improved performance while preserving the detailed texture of the original image.

2. Network Model

3. Structure-sensitive Multi-scale Deep Neural Network for Low-Dose CT Denoising

1. Summary

Computed tomography (CT) is a popular medical imaging method in clinical applications. Meanwhile, the dose of X-ray radiation associated with CT scans has drawn public attention due to its potential risk to patients. Over the past few years, efforts have been devoted to developing low-dose CT (LDCT) methods. However, reducing the radiation dose compromises the signal-to-noise ratio (SNR), resulting in strong noise and artifacts that degrade CT image quality. In this paper, we propose a novel 3D noise reduction method, called Structure-Sensitive Multi-Scale Generative Adversarial Network (SMGAN), to improve LDCT image quality. Specifically, we incorporate three-dimensional (3D) volume information to improve image quality. Furthermore, different loss functions for training denoising models are investigated. Experiments show that the method can effectively preserve the structural and texture information from normal dose CT (NDCT) images and significantly suppress noise and artifacts. Qualitative visual evaluation by three experienced radiologists showed that the method retrieves more detailed information and outperforms competing methods.

2. Network structure

4. Deep Convolutional Framelet Denosing for Low-Dose CT via Wavelet Residual Network

1. Summary

Model-based iterative reconstruction algorithms for low-dose X-ray computed tomography (CT) are computationally expensive. To address this issue, we recently proposed a deep convolutional neural network (CNN) for low-dose X-ray CT and won the second place in the 2016 AAPM Low-dose CT Grand Prix. However, some textures are not fully restored. To address this issue, here we propose a new wavelet frame-based denoising algorithm based on wavelet residual networks, which combines the expressive power of deep learning with the performance guarantees of wavelet frame-based denoising algorithms. The new algorithm is inspired by the recent interpretation of deep CNN s as cascaded convolutional framework signal representations. Extensive experimental results confirm that the proposed network has significantly improved performance and preserves the detailed texture of the original image.

2. Network structure


Posted by Rick Corbett on Sat, 07 May 2022 05:51:14 +0300