A hybrid CNN-adjoint optimization framework for reconstruction of viscoelastic tissue properties in magnetic resonance elastography

Anwesa Dey; Johann Rudi; Elena Cherkaev

math.OCOct 2, 2026

A hybrid CNN-adjoint optimization framework for reconstruction of viscoelastic tissue properties in magnetic resonance elastography

Authors: Anwesa Dey, Johann Rudi, Elena Cherkaev

Organizations: Department of Mathematics, The University of Utah, address USA. · Department of Mathematics, Virginia Tech, address USA. · Department of Mathematics, The University of Utah, address, address, USA.

Abstract

Magnetic resonance elastography (MRE) is a noninvasive imaging modality for quantifying the viscoelastic properties of soft tissues from shear wave propagation. Recovering the complex-valued shear modulus from measured displacement fields leads to a severely ill-posed inverse problem, particularly in the presence of noise and limited boundary excitations. We investigate adjoint-based optimization, convolutional neural network (CNN) reconstruction, and a hybrid framework combining both approaches. The forward model is based on a scalar form of the modified stationary Stokes system with a complex shear modulus. We establish well-posedness of the forward problem, existence of minimizers, and first-order optimality conditions for the adjoint-based formulation, and implement a nonlinear conjugate-gradient method with Armijo line search. While PDE-constrained optimization can accurately refine coefficient reconstructions, its performance depends strongly on initialization. We therefore construct a two-dimensional CNN that maps complex-valued displacement measurements to spatially varying complex shear modulus fields and provides rapid, informative initial reconstructions. The proposed hybrid method uses the CNN reconstruction to initialize the adjoint-based optimization, yielding faster convergence and improved accuracy. The CNN is trained on coefficient fields containing individual perturbations and tested on both individual and previously unseen combined configurations. Numerical experiments demonstrate that the CNN generalizes to these more challenging configurations, while subsequent PDE-constrained optimization further refines the reconstructed coefficient. These results demonstrate the potential of combining data-driven initialization with physics-based optimization for efficient and accurate MRE reconstruction.

Figures & tables

Figure 1: Representative complex-valued shear modulus $A(x,y)$ containing Gaussian perturbations in its real and imaginary components.

Figure 2: Real and imaginary parts of the displacement field $u(x,y)$ obtained by solving the forward PDE for the Gaussian shear modulus shown in Figure 1 .

Figure 3: Representative complex-valued shear modulus $A(x,y)$ containing mollified elliptical disc-shaped perturbations in its real and imaginary components.

Figure 4: Real and imaginary parts of the displacement field $u(x,y)$ obtained by solving the forward PDE for the mollified elliptical disc-shaped shear modulus shown in Figure 3 .

Stage	Operation	Output size
Input	–	$48\times 48\times 2$
Encoder 1	Conv2D $(3\times 3)$ , 32 filters, ReLU	$48\times 48\times 32$
Downsampling 1	Conv2D $(3\times 3)$ , 32 filters, stride 2, ReLU	$24\times 24\times 32$
Encoder 2	Conv2D $(3\times 3)$ , 64 filters, ReLU	$24\times 24\times 64$
Downsampling 2	Conv2D $(3\times 3)$ , 64 filters, stride 2, ReLU	$12\times 12\times 64$
Latent representation	2 Conv2D $(3\times 3)$ , 128 filters, ReLU	$12\times 12\times 128$

Table 1: Architecture of the 2D convolutional neural network used to generate the initial coefficient reconstruction.

	Mean	Std.	Median	Min.	Max.
$\operatorname{Re}(A)$	$6.4693\times 10^{-2}$	$3.0304\times 10^{-2}$	$5.8194\times 10^{-2}$	$1.4715\times 10^{-2}$	$2.1719\times 10^{-1}$
$\operatorname{Im}(A)$	$6.6093\times 10^{-2}$	$2.7720\times 10^{-2}$	$6.0846\times 10^{-2}$	$1.8937\times 10^{-2}$	$1.9950\times 10^{-1}$
$A$	$9.4681\times 10^{-2}$	$3.5716\times 10^{-2}$	$8.8932\times 10^{-2}$	$3.1948\times 10^{-2}$	$2.6949\times 10^{-1}$

Table 2: CNN reconstruction errors over the

500

test samples in Experiment I.

Panel of Figure 5 — Figure 5: Distribution of the CNN reconstruction errors over all $500$ test samples in Experiment I. The test dataset contains both Gaussian and mollified-ellipse coefficient fields.

Figure 6: Real and imaginary components of the internal displacement field $u_{d}$ corresponding to a representative complex-valued mollified-ellipse perturbation. These displacement data form the two input channels to the CNN.

Panel of Figure 7 — Figure 7: Representative CNN prediction for a complex-valued mollified-ellipse test sample from Experiment I. The top row compares the true and CNN-predicted real components of $A(x,y)$ , while the bottom row compares the corresponding imaginary components.

Figure 8: Real and imaginary components of the internal displacement field $u_{d}$ corresponding to a representative complex-valued Gaussian perturbation. These displacement data are supplied to the CNN as the two input channels for predicting the corresponding complex-valued coefficient field.

Panel of Figure 9 — Figure 9: Representative CNN prediction for a complex-valued Gaussian test sample from Experiment I. The top row compares the true and CNN-predicted real components of $A(x,y)$ , while the bottom row compares the corresponding imaginary components. The CNN prediction is subsequently used as the initial guess for the PDE-constrained optimization stage of the hybrid reconstruction framework.

Figure 10: Real component of the representative Gaussian test sample shown using separate color scales for the true coefficient and the CNN prediction. The individually scaled CNN prediction reveals the weak spatial variations that are not readily visible when the common color scale in Figure 9 is used.

Parameter	mollified elliptical disc	Gaussian
Data-fidelity weight $\alpha$	20	20
Regularization parameter $\beta$	$1\times 10^{-7}$	$5\times 10^{-6}$
Maximum NCG iterations $k_{\max}$	1000	2000
Initial line-search step size	1	1
Armijo parameter	$1\times 10^{-4}$	$1\times 10^{-4}$
Backtracking reduction factor	0.8	0.8

Table 3: Optimization parameters used for the representative mollified-ellipse and Gaussian test samples in Experiment I.

Figure 11: End-to-end hybrid reconstruction for a representative complex-valued mollified-ellipse test sample from Experiment I. The left column shows the true coefficient, the middle column shows the CNN-generated initial estimate, and the right column shows the coefficient reconstructed after PDE-constrained optimization. The top and bottom rows correspond to the real and imaginary components of $A(x,y)$ , respectively.

Figure 12: End-to-end hybrid reconstruction for a representative complex-valued Gaussian test sample from Experiment I. The left column shows the true coefficient, the middle column shows the CNN-generated initial estimate, and the right column shows the coefficient reconstructed after PDE-constrained optimization. The top and bottom rows correspond to the real and imaginary components of $A(x,y)$ , respectively.

	CNN initial estimate	Hybrid reconstruction	Error decrease (%)
$\\|\operatorname{Re}(A)-\operatorname{Re}(A_{d})\\|_{L^{2}}$	$1.005910\times 10^{-1}$	$3.075439\times 10^{-2}$	$69.43$
$\\|\operatorname{Im}(A)-\operatorname{Im}(A_{d})\\|_{L^{2}}$	$1.365835\times 10^{-1}$	$2.752220\times 10^{-2}$	$79.85$
$\\|A-A_{d}\\|_{L^{2}}$	$1.696279\times 10^{-1}$	$4.127111\times 10^{-2}$	$75.67$

Table 4: Reconstruction errors for the representative mollified-elliptical disc-shaped test sample before and after the hybrid reconstruction framework.

	CNN initial estimate	Hybrid reconstruction	Error decrease (%)
$\\|\operatorname{Re}(A)-\operatorname{Re}(A_{d})\\|_{L^{2}}$	$1.2582\times 10^{-1}$	$7.3615\times 10^{-2}$	$41.49$
$\\|\operatorname{Im}(A)-\operatorname{Im}(A_{d})\\|_{L^{2}}$	$4.3595\times 10^{-2}$	$4.1262\times 10^{-2}$	$5.35$
$\\|A-A_{d}\\|_{L^{2}}$	$1.3316\times 10^{-1}$	$8.4390\times 10^{-2}$	$36.62$

Table 5: Reconstruction errors for the representative Gaussian test sample before and after the hybrid reconstruction framework.

Panel of Figure 13 — Figure 13: Distribution of the CNN reconstruction errors over all $500$ test samples using internal displacement data contaminated with $1\%$ relative noise. The CNN is trained using noiseless displacement data and is applied directly to the noisy measurements without retraining.

	Mean	Std.	Median	Min.	Max.
$\operatorname{Re}(A)$	$7.1454\times 10^{-2}$	$3.2029\times 10^{-2}$	$6.5777\times 10^{-2}$	$1.5458\times 10^{-2}$	$2.3168\times 10^{-1}$
$\operatorname{Im}(A)$	$7.1907\times 10^{-2}$	$2.8521\times 10^{-2}$	$6.7771\times 10^{-2}$	$2.0490\times 10^{-2}$	$1.9553\times 10^{-1}$
$A$	$1.0377\times 10^{-1}$	$3.6707\times 10^{-2}$	$9.7522\times 10^{-2}$	$4.2700\times 10^{-2}$	$2.9684\times 10^{-1}$

Table 6: Summary statistics of the CNN reconstruction errors for the

1\%

noisy test data.

Parameter	Gaussian	mollified elliptical disc
Data-fidelity parameter $\alpha$	$20$	$20$
Regularization parameter $\beta$	$1\times 10^{-5}$	$1\times 10^{-6}$
Maximum number of iterations $k_{\max}$	$200$	$150$
Initial line-search step size	$1$	$1$
Armijo parameter	$10^{-4}$	$10^{-4}$
Step-size reduction factor	$0.8$	$0.8$

Table 7: Optimization parameters used for the reconstruction of the Gaussian and mollified-ellipse inclusions with

1\%

relative noise.

Reconstruction error	CNN initialization	Hybrid reconstruction	Error decrease (%)
$\\|\operatorname{Re}(A)-\operatorname{Re}(A_{d})\\|_{L^{2}(\Omega)}$	$6.291202\times 10^{-2}$	$6.229779\times 10^{-2}$	$0.98\%$
$\\|\operatorname{Im}(A)-\operatorname{Im}(A_{d})\\|_{L^{2}(\Omega)}$	$5.586334\times 10^{-2}$	$5.436543\times 10^{-2}$	$2.68\%$
$\\|A-A_{d}\\|_{L^{2}(\Omega)}$	$8.413462\times 10^{-2}$	$8.268383\times 10^{-2}$	$1.72\%$

Table 8: Reconstruction errors for the representative Gaussian coefficient using displacement data contaminated with

1\%

relative noise.

Figure 15: Reconstruction of a representative mollified-ellipse coefficient from internal displacement data contaminated with $1\%$ relative noise. The left, middle, and right columns show the true coefficient, the CNN-generated initial estimate, and the final hybrid reconstruction, respectively. The top and bottom rows correspond to the real and imaginary components of $A(x,y)$ .

Reconstruction error	CNN initialization	Hybrid reconstruction	Error decrease
$\\|\operatorname{Re}(A)-\operatorname{Re}(A_{d})\\|_{L^{2}(\Omega)}$	$7.901710\times 10^{-2}$	$7.279785\times 10^{-2}$	$7.87\%$
$\\|\operatorname{Im}(A)-\operatorname{Im}(A_{d})\\|_{L^{2}(\Omega)}$	$1.040895\times 10^{-1}$	$6.924087\times 10^{-2}$	$33.48\%$
$\\|A-A_{d}\\|_{L^{2}(\Omega)}$	$1.306841\times 10^{-1}$	$1.004680\times 10^{-1}$	$23.12\%$

Table 9: Reconstruction errors for the representative mollified-ellipstical disc coefficient using displacement data contaminated with

1\%

relative noise.

	Mean	Std.	Median	Min.	Max.
$\operatorname{Re}(A)$	$1.279967\times 10^{-1}$	$4.694296\times 10^{-2}$	$1.202852\times 10^{-1}$	$4.829479\times 10^{-2}$	$3.416799\times 10^{-1}$
$\operatorname{Im}(A)$	$1.290662\times 10^{-1}$	$4.606712\times 10^{-2}$	$1.234582\times 10^{-1}$	$4.722147\times 10^{-2}$	$3.733799\times 10^{-1}$
$A$	$1.852892\times 10^{-1}$	$5.509064\times 10^{-2}$	$1.760132\times 10^{-1}$	$8.382436\times 10^{-2}$	$4.479576\times 10^{-1}$

Table 10: CNN reconstruction errors over the

500

combined Gaussian and mollified-ellipse test samples in Experiment II.

Panel of Figure 16 — Figure 16: Distribution of the CNN reconstruction errors over all $500$ test samples in Experiment II. Each real and imaginary component of the coefficient contains both a Gaussian and a mollified-ellipse perturbation.

Figure 17: Real and imaginary components of the internal displacement field $u_{d}$ corresponding to a representative test coefficient containing combined Gaussian and mollified-ellipse perturbations. These displacement data form the two input channels to the CNN.

Panel of Figure 18 — Figure 18: Representative CNN prediction for a combined Gaussian and mollified-ellipse test sample from Experiment II. The top row compares the true and CNN-predicted real components of $A(x,y)$ , while the bottom row compares the corresponding imaginary components. Both components contain one Gaussian and one independently generated mollified elliptical disc.

Parameter	Value
Data-fidelity weight $\alpha$	20
Regularization parameter $\beta$	$1\times 10^{-7}$
Maximum NCG iterations $k_{\max}$	2500
Initial line-search step size	1
Armijo parameter	$1\times 10^{-4}$
Backtracking reduction factor	0.8

Table 11: Optimization parameters used throughout Experiment II.

Figure 19: End-to-end hybrid reconstruction for a representative combined Gaussian and mollified-ellipse test sample from Experiment II. The left column shows the true coefficient, the middle column shows the CNN-generated initial estimate, and the right column shows the coefficient reconstructed after PDE-constrained optimization. The top and bottom rows correspond to the real and imaginary components of $A(x,y)$ , respectively.

	CNN initial estimate	Hybrid reconstruction	Error decrease (%)
$\\|\operatorname{Re}(A)-\operatorname{Re}(A_{d})\\|_{L^{2}}$	$1.097275\times 10^{-1}$	$5.953433\times 10^{-2}$	$45.74$
$\\|\operatorname{Im}(A)-\operatorname{Im}(A_{d})\\|_{L^{2}}$	$7.435254\times 10^{-2}$	$4.500365\times 10^{-2}$	$39.47$
$\\|A-A_{d}\\|_{L^{2}}$	$1.325459\times 10^{-1}$	$7.463019\times 10^{-2}$	$43.70$

Table 12: Reconstruction errors for the representative combined Gaussian and mollified-ellipse test sample before and after the hybrid reconstruction framework.

Figure 20: End-to-end hybrid reconstruction for a representative test sample with spatially overlapping Gaussian and mollified-ellipse perturbations. The left column shows the true coefficient, the middle column shows the CNN-generated initial estimate, and the right column shows the coefficient obtained after PDE-constrained optimization. The top and bottom rows correspond to the real and imaginary components of $A(x,y)$ , respectively. This example provides a more challenging generalization test because the two inclusion types occupy overlapping spatial regions, while the CNN was trained only on the individual shape families.

	CNN initial estimate	Hybrid reconstruction	Error decrease (%)
$\\|\operatorname{Re}(A)-\operatorname{Re}(A_{d})\\|_{L^{2}}$	$8.751159\times 10^{-2}$	$5.142576\times 10^{-2}$	$41.24$
$\\|\operatorname{Im}(A)-\operatorname{Im}(A_{d})\\|_{L^{2}}$	$1.562200\times 10^{-1}$	$1.140009\times 10^{-1}$	$27.03$
$\\|A-A_{d}\\|_{L^{2}}$	$1.790614\times 10^{-1}$	$1.250632\times 10^{-1}$	$30.16$

Table 13: Reconstruction errors for the representative test sample with overlapping Gaussian and mollified-ellipse perturbations.

Panel of Figure 21 — Figure 21: Distribution of the CNN reconstruction errors over all $500$ combined-shape test samples using internal displacement data contaminated with $1\%$ relative noise. The CNN is trained using noiseless individual Gaussian and mollified-ellipse samples and is applied to the noisy combined-shape measurements without retraining.

Figure 22: Reconstruction of a representative combined-shape coefficient from internal displacement data contaminated with $1\%$ relative noise. The real component contains spatially separated Gaussian and mollified-ellipse inclusions, whereas the inclusions partially overlap in the imaginary component. The left, middle, and right columns show the true coefficient, the CNN-generated initial estimate, and the final hybrid reconstruction, respectively. The top and bottom rows correspond to the real and imaginary components of $A(x,y)$ .

Reconstruction error	CNN initialization	Hybrid reconstruction	Error decrease (%)
$\\|\operatorname{Re}(A)-\operatorname{Re}(A_{d})\\|_{L^{2}(\Omega)}$	$1.229673\times 10^{-1}$	$8.524580\times 10^{-2}$	$30.68$
$\\|\operatorname{Im}(A)-\operatorname{Im}(A_{d})\\|_{L^{2}(\Omega)}$	$1.193186\times 10^{-1}$	$6.553128\times 10^{-2}$	$45.08$
$\\|A-A_{d}\\|_{L^{2}(\Omega)}$	$1.713414\times 10^{-1}$	$1.075230\times 10^{-1}$	$37.24$

Table 14: Reconstruction errors for the representative combined-shape coefficient using displacement data contaminated with

1\%

relative noise.

Explore similar work

Jan 17, 2026

Accelerated MR Elastography Using Learned Neural Network Representation

To develop a deep-learning method for achieving fast high-resolution MR elastography from highly undersampled data without the need of high-quality training dataset. We first framed the deep neural network representation as a nonlinear extension of the linear subspace model, then used it to represent and reconstruct MRE image repetitions from undersampled k-space data. The network weights were learned using a multi-level k-space consistent loss. To further enhance reconstruction quality, phase-contrast specific magnitude and phase priors were incorporated, including the similarity of anatomical structures and smoothness of wave-induced harmonic displacement. Experiments were conducted using both 3D gradient-echo spiral and multi-slice spin-echo spiral MRE datasets. Compared to the conventional linear subspace-based approaches, the nonlinear network representation method was able to produce superior image reconstruction with suppressed noise and artifacts from a single in-plane spiral arm per MRE repetition (e.g., 2mm isotropic resolution in 1 min with a total R=10), yielding comparable stiffness estimation to the fully sampled data. This work demonstrated the feasibility of using deep network representations to model and reconstruct MRE images from highly-undersampled data, a nonlinear extension of the subspace-based approaches.

Xi Peng

Departments of Radiology, Biomedical Engineering, Electronic and Computer Engineering, and Mechanical Engineering, University of Iowa, Iowa City, IA 52246 · Department of Radiology, University of Iowa, Iowa City, IA, 52246

Representation Learning Accelerated MRI Reconstruction Magnetic Resonance Imaging

Ask PDF

Oct 1, 2026

Initial condition recovery in nonlinear damped viscous photoacoustic tomography using a convolutional neural network-guided gradient-free optimization framework

Photoacoustic tomography (PAT) is a hybrid imaging modality that combines high optical contrast with high ultrasonic resolution for biomedical imaging applications. In this work, we investigate the inverse problem of recovering the initial pressure distribution from boundary measurements in the presence of nonlinear acoustic propagation and viscous attenuation effects. To model these phenomena more accurately, we consider a nonlinear damped viscoelastic wave equation incorporating spatially varying sound speed, temporal attenuation, and nonlinear propagation mechanisms. We first establish the well-posedness of the corresponding forward problem using a Galerkin approximation combined with energy estimates and a fixed-point argument. For the inverse problem, we derive existence, uniqueness, and local uniqueness results under suitable assumptions through a harmonic extension reduction, spectral Laplace transform techniques, and observability estimates. To numerically reconstruct the initial pressure field, we develop a hybrid reconstruction framework that combines a convolutional neural network (CNN) with a gradient-free optimization strategy based on the sequential quadratic Hamiltonian (SQH) method derived from Pontryagin's maximum principle. The CNN is used to generate an informative initial guess, while the SQH framework enforces the governing PDE dynamics during the reconstruction process. Numerical experiments demonstrate that the proposed hybrid strategy significantly improves reconstruction quality, contrast, and robustness compared to standalone time-reversal and CNN-based approaches.

Madhu Gupta, Anwesa Dey, Prapti Tala +1

Department of Mathematics, Indian Institute of Technology, Gandhinagar, Gujarat, India. · Department of Mathematics, University of Utah, Salt Lake City, Utah, UT, USA. · Department of Mathematics, The University of Texas at Arlington, Arlington, TX, USA.

Medical Imaging PDE Inverse Problems Inverse Problems

Ask PDF

Jun 9, 2026

Non-linear mechanical field reconstruction coupling recurrent neural networks with physics-informed graph neural networks

Reconstructing local stress fields in heterogeneous microstructures under non-linear, history-dependent loading remains a major computational bottleneck in multi-scale simulations. We propose a coupled LSTM-GNN framework that links the temporal and spatial aspects of local stress field reconstruction. A Long Short-Term Memory network encodes macroscopic stress-strain sequences into a compact hidden state that captures the path-dependent constitutive response, while a physics-informed Graph Neural Network reconstructs the spatially-resolved stress field at each time step. We introduce a relative weighting strategy with linear warm-up to balance the data-driven reconstruction loss and a discrete divergence-based equilibrium penalty. This resolves the scale mismatch that prevents fixed-weight formulations from converging in the elasto-plastic regime. The model is trained on 10,000 non-proportional loading paths applied to a periodic plate-with-a-hole microstructure and von Mises elasto-plasticity. The model achieves three orders of magnitude speedup over finite element simulations and generalizes to loading sequences twice the training length, with 1.9% cumulative error. Because the graph relies on mesh connectivity instead of the specific element type, one trained surrogate can be applied directly without retraining to meshes with different element types and to both coarser and finer resolutions, while in all cases reproducing the high-fidelity quad-element FE field used during training. Indeed, the message passing characteristics inherent to GNN and MeshGraphNet architecture render the model mesh-agnostic. Analysis of the LSTM hidden states suggests a low-dimensional structure related to the internal state variables of the constitutive model.

Manuel Ricardo Guevara Garban, Yves Chemisky, Étienne Prulière +3

Univ. Bordeaux, CNRS, Bordeaux INP, I2M, UMR 5295, Talence, France · Univ. Bordeaux, CNRS, Bordeaux INP, LaBRI, UMR 5800, Talence, France · Institute of Mechanics and Fluid Dynamics, TU Bergakademie Freiberg, Freiberg, 09599, Germany +2

Neural Surrogate Modeling Structural Mechanics Physics-Informed ML Physics-Informed GNNs Spatiotemporal GNNs

Ask PDF