SGD-Net: Efficient Model-Based Deep Learning With Theoretical Guarantees

Liu, Jiaming; Sun, Yu; Gan, Weijie; Xu, Xiaojian; Wohlberg, Brendt; Kamilov, Ulugbek S.

doi:10.1109/tci.2021.3085534

Cited by 29 publications

(28 citation statements)

References 74 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Recently, convolutional neural networks (CNNs) have demonstrated impressive performance in MRI reconstruction. Strategies include unrolled networks [10,21,29,45], UNet-based networks [13,19], GAN-based networks [25,44], among others [20,40,50]. These learning methods have achieved state-of-the-art performance on public MRI challenge datasets [46].…”

Section: Traditional Methodsmentioning

confidence: 99%

End-to-End Sequential Sampling and Reconstruction for MRI

Yin¹,

Wu²,

Sun³

et al. 2021

Preprint

View full text Add to dashboard Cite

Section: Traditional Methodsmentioning

confidence: 99%

End-to-End Sequential Sampling and Reconstruction for MRI

Yin¹,

Wu²,

Sun³

et al. 2021

Preprint

View full text Add to dashboard Cite

“…ODER is compatible with any CNN architecture used to implement D θ . We use a tiny U-Net architecture [70] for ODER and the traditional RED (DEQ) [35]. We have added spectral normalization [74] to all the layers of CNN for stability (see the supplement for the numerical evaluation of the contractiveness of T θ on all three modalities).…”

Section: Numerical Evaluationmentioning

confidence: 99%

“…For reference we include several other well-known baseline methods, including TV [47], U-Net [4] and ISTA-Net + [23]. We also include the unfolded RED (Unfold) [70] and the traditional RED (Denoising) [11] to illustrate the improvements due to DEQ. TV is an iterative method that does not require training, while other methods are all DL-based with publicly available implementations.…”

Section: Numerical Evaluationmentioning

confidence: 99%

“…Many PnP/RED algorithms have been turned into DU architectures by parameterizing the operator D θ as a CNN with weights θ, truncating the PnP/RED algorithm to a fixed number of iterations, and training the corresponding architecture end-to-end in a supervised fashion. Recent work has explored strategies for reducing the memory and computational complexity of training DU architectures [69,70], However, a key bottleneck in DU training is the necessity to store the intermediate activation values required for computing the backpropagation updates, which fundamentally limits the number of unfolding layers one can practically use in large-scale applications. [33] is a recent method for training infinite-depth, weight-tied feedforward networks by analytically backpropagating through the fixed points using implicit differentiation.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Online Deep Equilibrium Learning for Regularization by Denoising

Liu¹,

Xu²,

Gan³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Plug-and-Play Priors (PnP) and Regularization by Denoising (RED) are widely-used frameworks for solving imaging inverse problems by computing fixed-points of operators combining physical measurement models and learned image priors. While traditional PnP/RED formulations have focused on priors specified using image denoisers, there is a growing interest in learning PnP/RED priors that are end-to-end optimal. The recent Deep Equilibrium Models (DEQ) framework has enabled memory-efficient end-to-end learning of PnP/RED priors by implicitly differentiating through the fixed-point equations without storing intermediate activation values. However, the dependence of the computational/memory complexity of the measurement models in PnP/RED on the total number of measurements leaves DEQ impractical for many imaging applications. We propose ODER as a new strategy for improving the efficiency of DEQ through stochastic approximations of the measurement models. We theoretically analyze ODER giving insights into its convergence and ability to approximate the traditional DEQ approach. Our numerical results suggest the potential improvements in training/testing complexity due to ODER on three distinct imaging applications.

show abstract

“…The recent publication [41] has reviewed PnP/RED in the context of image reconstruction for MRI. Deep unrolling is another widely-used strategy inspired by LISTA [42], where the iterations of a regularized optimization are interpreted as layers of a CNN and trained in an end-to-end fashion [31][32][33][42][43][44][45].…”

Section: Imaging Inverse Problemsmentioning

confidence: 99%

Deformation-Compensated Learning for Image Reconstruction without Ground Truth

Gan¹,

Sun²,

Eldeniz³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

Deep neural networks for medical image reconstruction are traditionally trained using high-quality ground-truth images as training targets. Recent work on Noise2Noise (N2N) has shown the potential of using multiple noisy measurements of the same object as an alternative to having a ground truth. However, existing N2N-based methods cannot exploit information from various motion states, limiting their ability to learn on moving objects. This paper addresses this issue by proposing a novel motion-compensated deep image reconstruction (MoDIR) method that can use information from several unregistered and noisy measurements for training. MoDIR deals with object motion by including a deep registration module jointly trained with the deep reconstruction network without any ground-truth supervision. We validate MoDIR on both simulated and experimentally collected magnetic resonance imaging (MRI) data and show that it significantly improves imaging quality.

show abstract

SGD-Net: Efficient Model-Based Deep Learning With Theoretical Guarantees

Cited by 29 publications

References 74 publications

End-to-End Sequential Sampling and Reconstruction for MRI

End-to-End Sequential Sampling and Reconstruction for MRI

Online Deep Equilibrium Learning for Regularization by Denoising

Deformation-Compensated Learning for Image Reconstruction without Ground Truth

Contact Info

Product

Resources

About