Convolutional Neural Networks for Inverse Problems in Imaging: A Review
read more
Citations
An overview of deep learning in medical imaging focusing on MRI
An overview of deep learning in medical imaging focusing on MRI
Machine learning for data-driven discovery in solid Earth geoscience
Deep learning on image denoising: An overview.
Solving inverse problems using data-driven models
References
Deep learning
Generative Adversarial Nets
Dropout: a simple way to prevent neural networks from overfitting
ImageNet classification with deep convolutional neural networks
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Related Papers (5)
Frequently Asked Questions (13)
Q2. What is the concrete perspective on CNNs?
The most concrete perspective on CNNs as generalizations of established algorithms comes from the idea of unrolling, which the authors discussed in the section “Network Architecture.”
Q3. What is the importance of patch size in a training set?
The patch size also has important ramifications for the performance of the network and is linked to its architecture, with larger filters and deeper networks requiring larger training patches [17].
Q4. What is the emerging paradigm for learning from sparse measurements?
The emerging paradigm is to learn to reconstruct from sparse measurements, using reconstructions from fully sampled measurements to train.
Q5. What is the way to avoid a small set of images dominating the error during training?
To avoid a small set of images dominating the error during training, it is best to scale the dynamic range of the training set [23], [27].
Q6. What can be done to improve the confidence in the results of a CNN?
demonstrating generalization between data sets (where the network learns on one data set, but is evaluated on another) can help improve confidence in the results by showing that the performance of the network is not dependent on some systematic bias of the data set.
Q7. What are the main areas of research that have been applied to inverse problems?
CNNs have so far been applied mostly to inverse problems where the measurements take the form of an image and the measurement model is simple, and less so for CT and MRI, which have relatively more complicated models.
Q8. Why is this critique of CNNs important?
While this critique can be made of any approach to inverse problems, it is especially relevant for CNNs because they are often treated as black boxes and because the reconstructions they generate are plausible-looking by design, hiding areas where the algorithm is less sure of the result.
Q9. What is the method for correcting artifacts created by direct or iterative?
Most of the surveyed works involve using a CNN to correct the artifacts created by direct or iterative methods, where it remains an open question what is the best such prereconstruction method.
Q10. What is the way to compare the results of CNN-based denoising?
For comparison, one very recent denoising work [11] reported a 0.7-dB improvement on a similar experiment, which remains less than 1 dB better than contemporary non-CNN methods (including block-matching and 3-D filtering, which had remained the state of the art for years).
Q11. How did one CNN approach achieve an average of 30.5 dB?
As another point of reference, in 2012, one CNN approach [7] reported an average PSNR of 30.2 dB on a set of standard test images (Lena, peppers, etc.), less than 0.1 dB better than comparisons, and another [8] reported an average of 30.5 dB on the same experiment.
Q12. What is the way to build the inverse operator into the network architecture?
One creative approach is to build the inverse operator into the network architecture as in [22], where the network can compute inverse Fourier transforms.
Q13. What is the general perspective on CNNs?
A more general perspective is that nearly all state-of-the-art iterative reconstruction algorithms alternate between linear steps and pointwise nonlinear steps, so it follows that CNNs should be able to perform similarly well given appropriate training.