Equivariant Deep Equilibrium Models for Imaging Inverse Problems

Alexander Mehta, Ruangrawee Kitichotkul, Vivek K. Goyal, Julián TachellaarXiv 2025 · DOI · arXiv

Abstract

Deep learning has transformed the field of imaging inverse problems, yet most state-of-the-art models rely on supervised learning with paired ground-truth data—a luxury often unavailable in clinical or scientific settings. This paper introduces a novel framework that combines Deep Equilibrium Models (DEQs) with Equivariant Imaging (EI), a self-supervised approach leveraging physical symmetries. By formalizing a modular approach to implicit differentiation, the authors demonstrate that DEQs can be trained effectively without ground-truth images. The resulting models, EI-DEQ, outperform standard U-Nets and unrolled networks in CT and MRI reconstruction tasks, achieving performance levels comparable to supervised models while maintaining superior memory efficiency.

Introduction

In imaging science, inverse problems involve the recovery of a high-fidelity signal $x$ from corrupted or incomplete measurements $y$. While deep neural networks have set new benchmarks in this domain, their dependence on large-scale, supervised datasets remains a significant bottleneck. In many practical applications, such as low-dose CT or accelerated MRI, obtaining "clean" ground-truth images is either technically impossible or ethically problematic.

To address this, researchers have turned to self-supervised learning. One promising avenue is Equivariant Imaging (EI), which utilizes the inherent symmetries (e.g., rotations, translations) of physical signals to learn reconstruction priors from measurement data alone. Simultaneously, Deep Equilibrium Models (DEQs) have emerged as a powerful architectural paradigm. Unlike traditional networks that pass data through a fixed number of layers, DEQs find the fixed point of a single layer, effectively simulating an infinitely deep network with constant memory cost. This paper bridges these two innovations, providing the mathematical and architectural framework to train DEQs using self-supervised equivariant losses.

Methodology

Modular Implicit Differentiation

The primary challenge in training DEQs lies in backpropagating through the fixed-point solution. This typically requires Implicit Differentiation (ID). When combined with complex, multi-pass loss functions like those found in EI, the mathematical derivation of gradients becomes increasingly cumbersome.

The authors address this by proving a Modular Implicit Differentiation property (Proposition 1). They demonstrate that the gradient of a DEQ fixed point can be computed modularly, allowing researchers to integrate complex self-supervised losses without manually deriving new fixed-point equations for every term in the loss function. This simplifies the implementation of DEQs in sophisticated training pipelines.

The EI-DEQ Framework

The proposed framework integrates the EI loss—which enforces both measurement consistency and equivariance under group transformations—into the DEQ training loop. The authors adapt two specific DEQ templates for inverse problems:

  1. DE-Grad: Based on the Regularization by Denoising (RED) framework, optimized for Sparse-View CT.
  2. DE-Prox: Based on the Proximal Gradient Method (PGM), optimized for Accelerated MRI.

By solving for the actual equilibrium point rather than truncating the process at a fixed number of iterations (as seen in "unrolled" networks), these models achieve a more stable and theoretically grounded reconstruction.

Experimental Setup

The researchers evaluated the EI-DEQ framework on two distinct imaging modalities:

  • Sparse-View CT: Utilizing the CT100 dataset (100 images, $128 \times 128$ resolution). The task involved reconstructing images from only 50 projection angles, a significant reduction from standard protocols.
  • Accelerated MRI: Utilizing the fastMRI dataset (199 knee scan slices). The task required $8\times$ acceleration—meaning $87.5\%$ of the k-space data was missing—compounded by additive Gaussian noise.

Key Insights and Results

Performance Benchmarks

The EI-DEQ models consistently outperformed both U-Net and unrolled architectures in self-supervised settings. In MRI reconstruction, the EI-DEQ achieved a Peak Signal-to-Noise Ratio (PSNR) of 31.34 dB, significantly higher than the EI U-Net (28.02 dB) and the 7-layer EI Unrolled network (27.24 dB). Similar gains were observed in CT reconstruction, where the EI-DEQ reached 35.41 dB.

The Importance of Exact Gradients

A critical finding of the study was the necessity of exact gradients. While "Jacobian-free backpropagation" (JFB) is a popular shortcut to speed up DEQ training, the authors found it led to significantly degraded performance in self-supervised tasks. Using exact Implicit Differentiation was essential for the model to converge to high-quality solutions.

Self-Supervision vs. Supervision

Perhaps most impressively, the self-supervised EI-DEQs performed within 1.3 dB of their fully supervised counterparts. This suggests that the physical prior provided by equivariance is a remarkably potent substitute for ground-truth labels. The models appear to learn an "invariant prior"—a mathematical understanding of image structure that respects the physical symmetries of the world.

Conclusion and Future Work

This work demonstrates that Equivariant Deep Equilibrium Models offer a robust, memory-efficient, and high-performing solution for imaging inverse problems where ground-truth data is scarce. By providing a modular approach to implicit differentiation, the authors have lowered the barrier to using DEQs in complex self-supervised frameworks.

Future research will focus on the computational overhead of DEQs, as the iterative solvers (such as Anderson Acceleration) required to find fixed points can be slower than the forward pass of traditional networks. Additionally, the authors aim to further investigate the nature of the "implicit prior" learned by these models and scale the framework to larger, more diverse clinical datasets.

Summary updated

Abstract

Equivariant imaging (EI) enables training signal reconstruction models without requiring ground truth data by leveraging signal symmetries. Deep equilibrium models (DEQs) are a powerful class of neural networks where the output is a fixed point of a learned operator. However, training DEQs with complex EI losses requires implicit differentiation through fixed-point computations, whose implementation can be challenging. We show that backpropagation can be implemented modularly, simplifying training. Experiments demonstrate that DEQs trained with implicit differentiation outperform those trained with Jacobian-free backpropagation and other baseline methods. Additionally, we find evidence that EI-trained DEQs approximate the proximal map of an invariant prior.
BibTeX
@misc{mehta2025equivariant,
  author = {Alexander Mehta and Ruangrawee Kitichotkul and Vivek K. Goyal and Julián Tachella},
  title = {{Equivariant Deep Equilibrium Models for Imaging Inverse Problems}},
  year = {2025},
  eprint = {2511.18667},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.18667},
}
Figure