CT image denoising with deep learning

01. A deep convolutional neural network using directional wavelets for low-dose X-ray CT reconstruction (KAIST-net)

paper

Dataset

• AAPM-Mayo Clinic Low-Dose CT Grand Challenge (only abdominal CT images)

• 512x512, 10 patients, 5743 slices
• use a 55x55 patches

Model

• This method works on wavelet coefficients of low-dose CT images
• Network contains 24 convolution layers
  

---

02. Low-dose CT via Convolutional Neural Network

paper

Dataset

• TCIA(The Cancer Imaging Archive) normal-dose CT images.

• 256x256, 165 patients, 7015 slices.
• impose Poisson noise into normal-dose sinogram.
• use a 33x33 patches.

Model

• Network use only 3 conoluional layers (Conv - ReLU - Conv - ReLU - Conv).

---

03. Improving Low-Dose CT Image Using Residual Convolutional Network

paper

Dataset

• AAPM-Mayo Clinic Low-Dose CT Grand Challenge

• 512x512, 10 patients, 5080 slices
• use a 44x44 patches(2D), 44x44x24 patches(3D)

Model

• 2D residual convolution net
• 3D residual convolution net (take into account the spatial continuity of tissues)
  

---

04. CT Image Denoising with Perceptive Deep Neural Networks

paper

Dataset

• cadaver CT image dataset collected at Massachusetts General Hospital (MGH)

Model

• Compare the denoised output against the ground truth in another high-dimensional feature space (from VGG)
  

---

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

paper

Dataset

• NBIA(Natioanl Biomedical Imaging Archive) normal-dose CT images

• 256x256, 165 patients, 7015 slices
• adding Poisson noise into the sinogram simulated from the normal-dose images

• AAPM-Mayo Clinic Low-Dose CT Grand Challenge

• 512x512, 10 patients, 2378 slices
• use a 55x55 patches

Model

• Incoporated a deconvolution network and shortcut connections into a CNN model
  

---

06. Generative adversarial networks for noise reduction in low-dose CT

paper

Dataset

• Phantom CT scans

• An anthropomorphic thorax phantom (QRM anthropomorphic thorax phantom)
• voltage of 120 kVp. 50mAs(routine-dose), 10mAs(low-dose)

• Cardiac CT scan (28 patients)

• voltage of 120 kVp. 5060mAs(routine-dose), 1012mAs(low-dose)

Model

• Generator transforms the low-dose CT image into noise reduced image
• Discriminator determines whether the input is a real routine-dose image or not
  

---

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

paper

Dataset

• AAPM-Mayo Clinic Low-Dose CT Grand Challenge

• 512x512, 10 patients, 2378 slices
• use a 80x80x11 patches

Model

• (Part 1). Generator consist of eight 3D convolutional (Conv) layers
• (Part 2). Calculate patch-wise error between the 3D output and the 3D NDCT images
  
• (Part 3). Discriminator distinguishes between two images
  

---

08. Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss

paper

Dataset

• AAPM-Mayo Clinic Low-Dose CT Grand Challenge

• 512x512, 10 patients, 4000 slices
• use a 64x64 patches

Model

• GAN with Wasserstein distance
• (Part 2). Comparing the perceptual feature of a denoised output against those of the ground truth in an established feature space
  

---

09. Sharpness-aware Low Dose CT Denoising Using Conditional Generative Adversarial Network

paper

Dataset

• NBIA(Natioanl Biomedical Imaging Archive) normal-dose CT images

• 512x512, 239 slices
• adding Poisson + normally Gaussian noise
• use a 256x256 patches (sampled from the 4 corners and center)

• Deceased piglet CT

• voltage of 100 kVp. 300mAs(full-dose) ~ 15mAs(low-dose)

• Phantom CT scans

• voltage of 120 kVp. 300mAs(full-dose) ~ 15mAs(low-dose)

• Data Science Bowl 2017

• Detect lung cancer from LDCTs

Model

• Sharpness detection network : generate a similar sharpness map as closs as to real CT
  

---

10. 3D Convolutional Encoder-Decoder Network for Low-Dose CT via Transfer Learning from a 2D Trained Network

paper

Dataset

• AAPM-Mayo Clinic Low-Dose CT Grand Challenge

• 512x512, 10 patients
• use a 64x64 patches

Model

• Concatenation of feature-maps from the two sides of the conveying-path
• Learn the 2D model first, and use it to initialize the 3D network. This transfer learning shows much faster convergence and better performance
  

---

11. Cycle Consistent Adversarial Denoising Network for Multiphase Coronary CT Angiography

paper

Dataset

• 50 CT scans of mitral valve prolapse patients, and 50 CT scans of coronary artery disease patients

• use a 56x56 patches

• AAPM-Mayo Clinic Low-Dose CT Grand Challenge

Model

• In coronary CTA, the images at the low-dose and routine-dose phases do not match each other exactly due to the cardiac motion
• Two generator denotes the mapping form low-dose to routine-dose image and from routine-dose to low-dose image, two adversarial discriminators distinguish between input images and synthesized images from the generators
• Using cycle-consistent adversarial denoising network, learn the mapping between the low and routine dose cardiac phases