Source-linked AI summary
Neural Proximal Gradient Descent for Compressive Imaging
Morteza Mardani, Qingyun Sun, Shreyas Vasawanala, Vardan Papyan, Hatef Monajemi, John Pauly, David Donoho
TL;DR
The paper addresses fast, physically plausible image recovery from severely undersampled measurements when training data are limited. It unrolls proximal-gradient iterations into a recurrent ResNet that learns the proximal map, and reports improved MRI reconstruction quality and efficiency over non-recurrent deep networks and CS-WV methods. The analysis also examines contraction and convergence, while acknowledging that fully rigorous, interpretable contraction guarantees remain incomplete.
Problem
Ill-posed inverse imaging requires plausible, physically consistent, and fast reconstructions despite limited labeled training data.
Method
The method unrolls proximal-gradient iterations into a recurrent architecture whose ResNet models the proximal map and is trained end-to-end.
Results
The architecture improves MRI results over general deep networks by about 2dB SNR and over CS-WV schemes by 4dB SNR, with 100x lower reconstruction time than CS-WV.
Takeaways & Limitations
A small recurrent ResNet can effectively model the proximal and improve the quality and complexity of deep and conventional CS-MRI architectures.
Takeaways & Limitations
The paper provides only local convergence insight; deriving simple, interpretable contraction guarantees and stable training for large iteration numbers remain ongoing work.
Abstract
from arXiv · showhide
Recovering high-resolution images from limited sensory data typically leads to a serious ill-posed inverse problem, demanding inversion algorithms that effectively capture the prior information. Learning a good inverse mapping from training data faces severe challenges, including: (i) scarcity of training data; (ii) need for plausible reconstructions that are physically feasible; (iii) need for fast reconstruction, especially in real-time applications. We develop a successful system solving all these challenges, using as basic architecture the recurrent application of proximal gradient algorithm. We learn a proximal map that works well with real images based on residual networks. Contraction of the resulting map is analyzed, and incoherence conditions are investigated that drive the convergence of the iterates. Extensive experiments are carried out under different settings: (a) reconstructing abdominal MRI of pediatric patients from highly undersampled Fourier-space data and (b) superresolving natural face images. Our key findings include: 1. a recurrent ResNet with a single residual block unrolled from an iterative algorithm yields an effective proximal which accurately reveals MR image details. 2. Our architecture significantly outperforms conventional non-recurrent deep ResNets by 2dB SNR; it is also trained much more rapidly. 3. It outperforms state-of-the-art compressed-sensing Wavelet-based methods by 4dB SNR, with 100x speedups in reconstruction time.
1 Introduction
The paper targets ill-posed image recovery under limited data, requiring fast, physically consistent, and plausible reconstructions. It proposes recurrent proximal learning and reports strong gains over non-recurrent networks and compressed-sensing Wavelet methods.
- Motivation: Ill-posed image restoration requires regularization while meeting speed, physical-consistency, and limited-training-data constraints.These challenges arise across applications including MRI visualization and interactive superresolution.
- Prior work: Prior learned approaches either sacrifice data-fidelity guarantees or rely on sparse-recovery objectives without explicit measurement modeling and convergence guarantees.Post-processing networks are fast but comparable in accuracy to iterative compressed sensing, while earlier learned sparse-recovery methods omit the measurement model.
- Contributions: A recurrent ResNet with a single residual block performs better when repeated than a general deep network for MRI reconstruction.The architecture is based on a neural proximal gradient descent scheme and is evaluated on pediatric MRI and natural-face superresolution.
- Results: 2dB SNR is the reported advantage over general deep network schemes, with less training data and faster training.The comparison is stated for the recurrent ResNet architecture in MRI reconstruction.
- Results: 4dB SNR and 100x lower reconstruction time separate the proposed architecture from state-of-the-art CS-WV schemes.The reported comparison concerns MRI reconstruction.
2 Preliminaries and problem statement
The paper formulates recovery as estimating a plausible high-dimensional image from undersampled, possibly noisy linear measurements. It connects this task to regularized least squares and proximal-gradient iterations for incorporating image priors.
- Problem formulation: The measurement model is y = Φx* + v with m ≪ n, where the unknown image lies on an otherwise unspecified low-dimensional manifold.Training samples and corresponding possibly noisy observations are available, but no other manifold information is assumed.
- Problem formulation: The goal is to quickly recover a plausible image x* from a new undersampled observation y.The formulation covers restoration tasks such as MRI reconstruction and superresolution.
- Prior-based estimation: A regularized least-squares program incorporates an image prior through a parameterized regularizer ψ(·).The forward operator represents the acquisition physics or downsampling process, depending on the task.
- Proximal optimization: Proximal gradient applies a data-fidelity step followed by a proximal operator determined by the regularizer.For convex ψ, the proximal map is monotone and the fixed point coincides with the global optimum; ℓ1 regularization yields soft-thresholding in a transform domain.
3 Neural Proximal learning
The proposed architecture unrolls proximal-gradient iterations into a recurrent network that combines data consistency with a learned proximal map. The proximal is modeled by a neural network and trained end-to-end over repeated iterations.
- Design motivation: The design asks how to achieve rapid inference with affordable training for real-time image recovery.The recurrence is motivated by proximal-gradient iterations rather than by a standalone image-to-image mapping.
- Iterative architecture: Each iteration moves the current estimate toward data-consistent images before applying the proximal operator to produce the next estimate.The gradient step uses the measurement y and the forward operator Φ.
- Proximal modeling: A truncated recurrent network receives the measurement at every iteration and combines it with the previous output to model the proximal map.The proximal is interpreted as a projection toward visually plausible images or as a denoiser.
- Proximal modeling: The proximal network uses K layers with element-wise activations, typically convolutional weights, and ResNet or highway connections to reduce vanishing-gradient problems.At each iteration, the network maps z0 = xt to zK = xt+1.
- Training: The T-step reconstruction is the repeated composite map (Pψ ◦ g)^T applied to the initial linear estimate ΦHy, with parameters learned end-to-end.Training can use element-wise losses such as MSE, while a measurement-consistency term improves convergence for large T.
4 Contraction Analysis
The contraction analysis studies convergence of recurrent neural proximal-gradient iterates under noiseless measurements, representation error, and fixed-point assumptions. It derives sufficient contraction conditions, while identifying sparse perturbation behavior and de-biasing as practical or open considerations.
- Analysis setup: The inference analysis asks whether the recurrent iterates converge, how quickly they converge, and whether their limit matches the unknown image.The analysis is conducted under explicit assumptions to make these questions tractable.
- Assumptions: The analysis assumes noiseless measurements y = Φx∗ and that the true image is close to a fixed point of the learned proximal map.The fixed-point assumption informally treats images as lying on a manifold represented by Pψ and can be encouraged during training.
- Error recursion: The iterate error is decomposed into terms involving the fixed proximal map, perturbation map, measurement operator, representation error, and the unknown image.This decomposition is the basis for bounding the propagation of the error across iterations.
- Contraction conditions: A sufficient condition for asymptotic linear convergence is lim sup as t→∞ of (η_t^1 + η_t^2) being less than 1 when the representation error is approximately zero.The contraction parameters are defined separately for the fixed map and its perturbation.
- Contraction conditions: For non-negligible representation error, the analysis instead gives a sufficient condition for convergence within a ν-ball around x∗.The resulting guarantee is approximate rather than exact because the representation error contributes to the iterate bound.
- Scope and open questions: The authors conjecture that perturbation masks become highly sparse after enough iterations, but deriving guarantees from interpretable network-parameter conditions remains future work.Empirical observations of contraction parameters are provided, while the stronger theoretical analysis is deferred.
- Practical refinement: After convergence, fixing the layer masks and replacing the proximal map with a linear map is proposed as a way to reduce bias from representation error.This de-biasing step is analogous to post-selection least-squares de-biasing in sparse linear regression.
5 Experiments
Experiments evaluate recurrent neural proximal gradient descent for undersampled pediatric MRI and face superresolution, comparing architecture depth, iteration count, conventional deep networks, and CS-WV. Results show that repeated small proximal networks improve reconstruction quality while reducing training and inference costs.
- Experimental scope: The experiments assess pediatric MRI reconstruction from undersampled k-space and 2× natural-face superresolution, with superresolution also used to examine contraction behavior.The study compares conventional deep architectures and CS-MRI, varies proximal depth and iteration count, and tests empirical contraction conditions.
- Architecture and training: A recurrent ResNet uses residual blocks composed of two 3 × 3 convolutional layers with 128 feature maps, followed by batch normalization and ReLU.Three additional 1 × 1 convolutional layers complete the proximal network.
- MRI setting: 20% of Fourier coefficients were collected for single-coil MRI using variable-density radial sampling that favors low-frequency k-space components.The dataset contains T1-weighted abdominal volumes from 350 pediatric patients, with 10K training slices and 1,280 test slices used for the architecture study.
- Iteration and depth: More than 2dB SNR gain for 1 RB and more than 3dB for 2 RBs is achieved by using three iterations instead of one.With one iteration, 10 RBs reaches SNR=24.33 while 5 RBs reaches SNR=24.15, showing limited benefit from simply deepening the network.
- Efficiency: 2–3 hours of training with shared weights and 10 iterations with 1 RB contrasts with around 10–12 hours for a deep ResNet with 10 RBs.One unrolled iteration with one RB takes 4 msec on the GPU, while inference time grows with the number of iterations.
- MRI comparisons: More than 4dB SNR gain over CS-WV is reported for the recurrent scheme, with sharper liver-vessel details and 100x reduced reconstruction computing time.Ten iterations with one RB produce more visible vessel details than a one-iteration, 10-RB ResNet; five iterations with two RBs outperform two iterations with five RBs qualitatively.
6 Conclusions
The paper develops neural proximal gradient descent with recurrent ResNets for compressed image recovery, and evaluates both reconstruction quality and convergence. It finds that small recurrent ResNets improve MRI recovery while the analysis of local convergence remains incomplete.
- A recurrent neural proximal gradient scheme models the proximal map with ResNets for recovering images from highly compressed measurements.The approach is evaluated on pediatric MRI reconstruction and natural-image superresolution, with contraction and local convergence studied empirically.
- The paper’s convergence analysis is local, and deriving simple, interpretable contraction conditions remains future work.The authors identify the distribution of activation masks as requiring extensive empirical evaluation.