Unsupervised Training of Denoisers for Low-Dose CT Reconstruction Without Full-Dose Ground Truth

Kim, Kwanyoung; Soltanayev, Shakarim; Chun, Se Young

doi:10.1109/jstsp.2020.3007326

Cited by 34 publications

(31 citation statements)

References 49 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This approach may over smoothed the edges as the data is compressed by the encoders and decoders. The PURE [42] method uses skip connections between the layers to form the residue and denoise the LDCT images. It uses modified U-Net network and full-dose CT data as ground truth images although noise is present in the images.…”

Section: A Analysis Of Proposed Augmented Noise Learning Methodsmentioning

confidence: 99%

“…In Fig. 14 (c) one can observe that the image obtained by PURE [42] still has a bit of noise present. Fig.…”

Section: Experiments and Evaluationmentioning

confidence: 95%

“…In order to denoise LDCT images, a tunable CycleGAN is proposed using only one generator that makes use of adaptive instance normalization (AdaIN) layers [41]. A deep neural network is recently trained in an unsupervised way using Poisson unbiased risk estimator (PURE) to denoise the LDCT image [42]. The RL is a promising method to learn the noise from the images, still, overfitting occurs as it starts learning the image content along with the noise [43], [44].…”

Section: Related Workmentioning

confidence: 99%

“…As shown in the figure, they are fixed to λ = 0.5 and µ = 1 as they show highest PSNR values and help avoiding minimax and overfitting problems. For implementing other comparative approaches, we use the optimal values of parameters as available in respective papers [22], [34], [35], [37]- [39], [41], [42], [58]- [61].…”

Section: Experiments and Evaluationmentioning

confidence: 99%

See 3 more Smart Citations

Augmented Noise Learning Framework for Enhancing Medical Image Denoising

2021

View full text Add to dashboard Cite

Deep learning attempts medical image denoising either by directly learning the noise present or via first learning the image content. We observe that residual learning (RL) often suffers from signal leakage while dictionary learning (DL) is prone to Gibbs (ringing) artifacts. In this paper, we propose an unsupervised noise learning framework that enhances denoising by augmenting the limitation of RL with the strength of DL and vice versa. To this end, we propose a ten-layer deep residue network (DRN) augmented with patch-based dictionaries. The input images are presented to patch-based DL to indirectly learn the noise via sparse representation while given to the DRN to directly learn the noise. An optimum noise characterization is captured by iterating DL/DRN network against proposed loss. The denoised images are obtained by subtracting the learned noise from available data. We show that augmented DRN effectively handles high-frequency regions to avoid Gibbs artifacts due to DL while augmented DL helps to reduce the overfitting due to RL. Comparative experiments with many state-of-the-arts on MRI and CT datasets (2D/3D) including low-dose CT (LDCT) are conducted on a GPU-based supercomputer. The ablation studies are carried out that demonstrate enhanced denoising performance with minimal signal leakage and least artifacts by proposed augmented approach.

show abstract

Section: A Analysis Of Proposed Augmented Noise Learning Methodsmentioning

confidence: 99%

“…In Fig. 14 (c) one can observe that the image obtained by PURE [42] still has a bit of noise present. Fig.…”

Section: Experiments and Evaluationmentioning

confidence: 95%

Section: Related Workmentioning

confidence: 99%

Section: Experiments and Evaluationmentioning

confidence: 99%

See 2 more Smart Citations

Augmented Noise Learning Framework for Enhancing Medical Image Denoising

2021

View full text Add to dashboard Cite

show abstract

“…major focus of research in the CT community [5]- [8]. In particular, the authors in [5], [6] proposed a CycleGAN approach [9] for low-dose CT denoising that trains the networks with unpaired LDCT and SDCT images.…”

Section: Introductionmentioning

confidence: 99%

Cycle-free CycleGAN using Invertible Generator for Unsupervised Low-Dose CT Denoising

Kwon¹,

Ye²

2021

Preprint

View full text Add to dashboard Cite

Recently, CycleGAN was shown to provide highperformance, ultra-fast denoising for low-dose X-ray computed tomography (CT) without the need for a paired training dataset. Although this was possible thanks to cycle consistency, CycleGAN requires two generators and two discriminators to enforce cycle consistency, demanding significant GPU resources and technical skills for training. A recent proposal of tunable CycleGAN with Adaptive Instance Normalization (AdaIN) alleviates the problem in part by using a single generator. However, two discriminators and an additional AdaIN code generator are still required for training. To solve this problem, here we present a novel cyclefree Cycle-GAN architecture, which consists of a single generator and a discriminator but still guarantees cycle consistency. The main innovation comes from the observation that the use of an invertible generator automatically fulfills the cycle consistency condition and eliminates the additional discriminator in the CycleGAN formulation. To make the invertible generator more effective, our network is implemented in the wavelet residual domain. Extensive experiments using various levels of low-dose CT images confirm that our method can significantly improve denoising performance using only 10% of learnable parameters and faster training time compared to the conventional CycleGAN.

show abstract

Learning low‐dose CT degradation from unpaired data with flow‐based model

Liang

Deng

et al. 2022

Medical Physics

View full text Add to dashboard Cite

Background There has been growing interest in low‐dose computed tomography (LDCT) for reducing the X‐ray radiation to patients. However, LDCT always suffers from complex noise in reconstructed images. Although deep learning‐based methods have shown their strong performance in LDCT denoising, most of them require a large number of paired training data of normal‐dose CT (NDCT) images and LDCT images, which are hard to acquire in the clinic. Lack of paired training data significantly undermines the practicability of supervised deep learning‐based methods. To alleviate this problem, unsupervised or weakly supervised deep learning‐based methods are required. Purpose We aimed to propose a method that achieves LDCT denoising without training pairs. Specifically, we first trained a neural network in a weakly supervised manner to simulate LDCT images from NDCT images. Then, simulated training pairs could be used for supervised deep denoising networks. Methods We proposed a weakly supervised method to learn the degradation of LDCT from unpaired LDCT and NDCT images. Concretely, LDCT and normal‐dose images were fed into one shared flow‐based model and projected to the latent space. Then, the degradation between low‐dose and normal‐dose images was modeled in the latent space. Finally, the model was trained by minimizing the negative log‐likelihood loss with no requirement of paired training data. After training, an NDCT image can be input to the trained flow‐based model to generate the corresponding LDCT image. The simulated image pairs of NDCT and LDCT can be further used to train supervised denoising neural networks for test. Results Our method achieved much better performance on LDCT image simulation compared with the most widely used image‐to‐image translation method, CycleGAN, according to the radial noise power spectrum. The simulated image pairs could be used for any supervised LDCT denoising neural networks. We validated the effectiveness of our generated image pairs on a classic convolutional neural network, REDCNN, and a novel transformer‐based model, TransCT. Our method achieved mean peak signal‐to‐noise ratio (PSNR) of 24.43dB, mean structural similarity (SSIM) of 0.785 on an abdomen CT dataset, mean PSNR of 33.88dB, mean SSIM of 0.797 on a chest CT dataset, which outperformed several traditional CT denoising methods, the same network trained by CycleGAN‐generated data, and a novel transfer learning method. Besides, our method was on par with the supervised networks in terms of visual effects. Conclusion We proposed a flow‐based method to learn LDCT degradation from only unpaired training data. It achieved impressive performance on LDCT synthesis. Next, we could train neural networks with the generated paired data for LDCT denoising. The denoising results are better than traditional and weakly supervised methods, comparable to supervised deep learning methods.

show abstract

Unsupervised Training of Denoisers for Low-Dose CT Reconstruction Without Full-Dose Ground Truth

Cited by 34 publications

References 49 publications

Augmented Noise Learning Framework for Enhancing Medical Image Denoising

Augmented Noise Learning Framework for Enhancing Medical Image Denoising

Cycle-free CycleGAN using Invertible Generator for Unsupervised Low-Dose CT Denoising

Learning low‐dose CT degradation from unpaired data with flow‐based model

Contact Info

Product

Resources

About