Skip to content

200202-200208-papers-reading.md

Latest commit

 

History

History
116 lines (56 loc) · 9.62 KB

200202-200208-papers-reading.md

File metadata and controls

116 lines (56 loc) · 9.62 KB

2000202-200208 论文阅读总结

[TOC]

论文列表

  1. Uncertainty Autoencoders: Learning Compressed Representations via Variational Information Maximization: UAE

  2. Hyperspectral Image Reconstruction Using a Deep Spatial-Spectral Prior: Hyperspectral Image

  3. Efficient Structurally-Strengthened Generative Adversarial Network for MRI Reconstruction: MRI

  4. Dynamic MRI using model‐based deep learning and SToRM priors: MoDL‐SToRM: MRI

  5. Spatio-Temporal Deep Learning-Based Undersampling Artefact Reduction for 2D Radial Cine MRI with Limited Training Data: MRI

  6. One-dimensional Deep Image Prior for Time Series Inverse Problems: DIP

  7. Sinogram interpolation for sparse-view micro-CT with deep learning neural network: CT

  8. An improved method for single image super-resolution based on deep learning: SR

  9. Deep MR Fingerprinting with total-variation and low-rank subspace priors

  10. Alternating Phase Projected Gradient Descent with Generative Priors for Solving Compressive Phase Retrieval: phase retrieval

  11. Deep learning for low-dose CT：CT

  12. Information-Theoretic Lower Bounds for Compressive Sensing with Generative Models：theory

  13. Deep Residual Dense U-Net for Resolution Enhancement in Accelerated MRI Acquisition: MRI

  14. GrappaNet: Combining Parallel Imaging with Deep Learning for Multi-Coil MRI Reconstruction: MRI

  15. Neumann Networks for Linear Inverse Problems in Imaging: inverse problem

  16. Regularizing linear inverse problems with convolutional neural networks：theory

  17. NETT Regularization for Compressed Sensing Photoacoustic Tomography: PAT

  18. Deep Learning for Inverse Problems: Bounds and Regularizers：regularization of deep learning

  19. A Very Deep Densely Connected Network for Compressed Sensing MRI:

  20. Learning Sub-Sampling and Signal Recovery with Applications in Ultrasound Imaging:

  21. SURE-TISTA: A Signal Recovery Network For Compressed Sensing: CS

  22. Error Resilient Deep Compressive Sensing:

  23. Low Shot Learning with Untrained Neural Networks for Imaging Inverse Problems: inverse problem

  24. A Theoretically Guaranteed Deep Optimization Framework for Robust Compressive Sensing MRI: MRI

  25. Accelerated Projection Reconstruction MR imaging using Deep Residual Learning: MRI

  26. Surfing: Iterative Optimization Over Incrementally Trained Deep Networks: optimization

  27. Non-Learning based Deep Parallel MRI Reconstruction (NLDpMRI): MRI

  28. Assessment of the generalization of learned image reconstruction and the potential for transfer learning: MRI

  29. Deep Decomposition Learning for Inverse Imaging Problems: inverse problem

  30. Real-time Cardiovascular MR with Spatio-temporal Artifact Suppression using Deep Learning - Proof of Concept in Congenital Heart Disease: MRI

  31. Robust contrast-transfer-function phase retrieval via flexible deep learning networks：phase retrieval

  32. DeepcomplexMRI: Exploiting deep residual network for fast parallel MR imaging with complex convolution: MRI

  33. Variational Deep Learning for Low-dose Computed Tomography: CT

  34. Learning from our neighbours: a novel approach on sinogram completion using bin-sharing and deep learning to reconstruct high quality 4DCBCT: CT

  35. Deep Compressed Sensing: CS

  36. pISTA-SENSE-ResNet for Parallel MRI Reconstruction: MRI

  37. λ-net: Reconstruct Hyperspectral Images from a Snapshot Measurement: hyper spectral image

  38. IFR-Net: Iterative Feature Refinement Network for Compressed Sensing MRI: MRI

  39. Learning Priors in High-frequency Domain for Inverse Imaging Reconstruction: inverse problem

  40. On the existence of stable and accurate neural networks for image reconstruction

  41. Insights into Learning-Based MRI Reconstruction: Overview

  42. Learning the Weight Matrix for Sparsity Averaging in Compressive Imaging: CS

  43. Learning to solve inverse problems using Wasserstein loss

  44. fastMRI: An Open Dataset and Benchmarks for Accelerated MRI: Dataset

Ideas from papers

  • 在 Spatio-temporal 的 3D 重建问题中，使用空间和时间两个方向的的 slice 作为输入，可以增加很多训练数据（Spatio-Temporal Deep Learning-Based Undersampling Artefact Reduction for 2D Radial Cine MRI with Limited Training Data）
  • 把重建信号分解成正交两个部分（Deep Decomposition Learning for Inverse Imaging Problems）

个人思考

  • 如何解释 DIP，也许网络结构起到的是一个超定方程的框架，DIP 的有效要从函数表示的角度入手，猜测 DIP 的参数就是在求解一个类似线性方程的系统，只不过是局部、非线性的。在每个局部都是一个超定方程（而且超定的程度很高），多个超定方程构成一个超定程度较低的方程），这样以来，对于随机的输入，总是可以找到一组解。为什么 DIP 一般比较快的生成低频部分或简单部分，应该就是因为 SGD 算法在求解超定方程时的特性吧，网络结构的正则项也许只是提供了求解系统的框架。重建信号本身的性质和超定方程求解过程中达成了某种一致性，才会让DIP的效果好。如果网络结构不同，对应不同的求解系统，那么对于需要重建的信号，效果可能就不一样。如果按照这样的解释，也可以用一些较为简单的模型来证明一些结论，也可以设计各种实验来验证，比如设计一些反常的信号来重建。这个思路是否可以用来解释 GAN 等其他方法？
  • 分片线性函数和 relu 的关系，激活值为0是不是某种函数表示的控制。任何一个1维的分片函数都可以用多层的带 relu 的线性函数来表示，每一层都可以增加分片的粒度。

讨论：

noise2noise 和 DIP 的关系

生成模型