[DL] SuperResolution in Magnetic Resonance Imaging

1. Super-resolution MRI through Deep Learning (Qing Lyu et al, 2018 )

1. Two SOTA NN for CT denoting and deblurring, transfering them for super-resolution MRI
2. Two-fold resolution enhancement

2. CT Super-resolution GAN Constrained by the Identical, Residual, and Cycle Learning Ensemble(GAN-CIRCLE)

• Reference of [1]

3. MRI Super-Resolution with Ensemble Learning and Complementary Priors ( Qing Lyu et al, 2019 )

1. Obtaining high-quality MR image is often challenging
2. Ensemble learning and DL framework for MR image Super Resolution
   1. First enlarged low resolution images using 5 commonly used super-resolution algs and obtained differentiallly enlarged image datasets with complementary priors
   2. Then, GAN is trained with each dataset to generate super-resolution MR images
   3. Finally, CNN is used for ensemble learning that synergies the outputs of GANs into the final MR SR images.
   4. According to results, the ensemble learning results outcome any one of GAN outputs

4. Efficient and Accurate MRI Super-Resolution Using a Generative Adversarial Network and 3D Multi-level Densely Connected Network ( Yuhua Chen, 2018, MICCAI )

1. 3d NN design - multi-level densely connected super-resolution network (mDCSRN) with GAN

2. The GAN promotes realistic output hardly distinguishable from original HR images

5. Brain MRI super resolution using 3D deep densely connected neural networks ( Yuhua Chen et al, 2018, ISBI )

1. Single Image Super resolution (SISR)
2. 3D Densely Connected Super Resolution Network ( DCSRN )
3. 4x resolution-reduced images | This model outperforms bicubic interpolation

6. Enhanced generative adversarial network for 3D brain MRI super-resolution ( Jiancong Wang et al, 2020, WACV : winter conference on Applications of CV )

1. Single Image Super resolution (SISR) on 3D images

2. GAN framework and developed a generator coupled w/ discriminator to tackle the task of 3D SISR on T1 brain MRI images

3. Developed a 3D memory-efficient residual-dense block generator ( MRDG ) that achieves SOTA performance in terms of PSNR, SSIM, NRMSE ( normalized MSE Error )

4. Pyramid pooling discriminator ( PPD ) to recover details on diff. size scales simultaneously

5. Introduced a model blending, simple and computational efficient method to balance btw image and texture quality in the final output

7. Image super-resolution using progressive generative adversarial networks for medical image analysis ( Dwarikanath Mahaptra et al, 2018, Computerized Medical Imaging and Graphics )

1. Proposed an image super-resolution method using progressive GAN ( P - GANs ) which takes an input a low resolution image and generate a high resolution image of desired scaling factor
2. Primary Contribution : proposing a multi-stage model where the output image quality of one stage is progressively improved in the next stage by using a triplet loss function
3. Triplet loss enables stepwise image quality improvement by using the output of the previous stage as the baseline
4. Results show that this multi-stage P-GAN outperforms competing methods and baseline GANs.

 

출처 : https://github.com/SWKoreaBME/paper_review/blob/master/Review/06.MRI%20super-resolution%20literature%20survey.md

SWKoreaBME/paper_review

* 본 설명은 상기 논문들의 Abstract 을 요약한 것이며 문장이나 단어가 겹치는 것이 있을 수 있습니다. 요약 및 설명은 SWKoreaBME/paper_review Github 에서 가져온 것입니다.