GENRE-CMR: Robust Cardiac MRI Reconstruction

• GENRE-CMR is a supervised GAN-based framework for accelerated cardiac MRI that fuses residual unrolled reconstruction with dual-domain processing.
• It integrates adversarial training, edge-aware loss, and statistical distribution alignment to preserve key anatomical details and generalize across imaging conditions.
• Empirical results on the CMRxRecon 2025 dataset demonstrate enhanced SSIM, PSNR, and reduced NMSE, particularly on unseen distributions.

GENRE-CMR is a supervised, GAN-based framework for accelerated cardiac MRI reconstruction that is designed to remain effective across heterogeneous acquisition domains rather than only on the narrow distribution seen during training. Introduced as “Generalizable Deep Learning for Diverse Multi-Domain Cardiac MRI Reconstruction,” it addresses the standard accelerated MRI inverse problem under undersampled k-space while targeting the practical generalization gap created by variation across imaging centers, scanner vendors, protocols, image contrasts, trajectories, anatomical coverage, and patient populations (Hamedani et al., 28 Aug 2025). In this formulation, “generalizable” denotes robustness to out-of-distribution CMR data not represented in the training distribution, and “multi-domain” denotes variation across imaging centers, scanner vendors, protocols, image contrasts, trajectories, anatomical coverage/views, and other acquisition conditions (Hamedani et al., 28 Aug 2025).

1. Definition and clinical-technical setting

GENRE-CMR operates on multi-coil undersampled k-space and reconstructs an image through a sequence of unrolled reconstruction stages (Hamedani et al., 28 Aug 2025). The framework is positioned against reconstruction approaches that either perform well but are not robust to domain shift, rely on heavy diffusion models or external natural-image priors, or do not jointly combine unrolled optimization, adversarial learning, and explicit domain alignment (Hamedani et al., 28 Aug 2025). Its stated objective is a single unified CMR reconstruction model that can handle diverse multi-domain data and maintain strong performance on unseen distributions (Hamedani et al., 28 Aug 2025).

The method is explicitly dual-domain. K-space is used for data fidelity, physical consistency, and the unrolled update; image space is used for learned refinement, SSIM-based reconstruction, and edge-aware supervision (Hamedani et al., 28 Aug 2025). It is also explicitly adversarial: a generator performs reconstruction using a residual deep unrolled architecture, and a discriminator is trained adversarially to encourage more realistic reconstructions (Hamedani et al., 28 Aug 2025).

This positioning places GENRE-CMR within a broader trajectory of generative and structure-aware CMR methods. Earlier work on post-reconstruction artifact correction had already shown that adversarial learning and edge-focused supervision could sharpen CMR outputs in image space, although in a different problem setting centered on motion artifact correction rather than accelerated reconstruction (Zhang et al., 2019). A plausible implication is that GENRE-CMR extends that broader pattern from post hoc enhancement to physics-constrained undersampled reconstruction while shifting the emphasis from single-domain restoration to multi-domain robustness.

2. Residual deep unrolled reconstruction framework

The central architectural novelty is a residual deep unrolled reconstruction framework (Hamedani et al., 28 Aug 2025). GENRE-CMR builds on an earlier 2024 model and modifies it by adding residual connections between consecutive cascaded modules so that feature information from shallow stages is preserved and reused by later-stage reconstructors (Hamedani et al., 28 Aug 2025). The authors argue that this improves convergence, hierarchical representation learning, and gradient flow (Hamedani et al., 28 Aug 2025).

The generator consists of 16 reconstructors, also described as cascades or unrolled steps (Hamedani et al., 28 Aug 2025). At each step, the current k-space estimate is transformed to image space, coil sensitivity information is used for coil combination and processing, a convolutional reconstruction subnetwork predicts a feature or refinement term, that prediction is added residually to the propagated feature from the previous stage, the refined image representation is mapped back into multi-coil k-space, and a data-consistency-like update enforces agreement with acquired samples (Hamedani et al., 28 Aug 2025). The algorithm is given as:

$$k^{(ACS)} \leftarrow \mathrm{Auto\_calibration\_signal}(k^{(0)})$$

$I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$

$$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$

Residual features are initialized as

$F^{(-1)} \leftarrow 0$

and for each unrolled step $t \in \{0,\dots,T-1\}$ with $T=16$,

$$I_{MC} \leftarrow \mathrm{FFT^{-1}}(k^{(t)}), \qquad I_{SC} \leftarrow S_m' \cdot I_{MC}$$

$$\hat{F}^{(t)} \leftarrow \mathrm{APUnet}^{(t)}(I_{SC}, P^{(t)})$$

$F^{(t)} \leftarrow \hat{F}^{(t)} + F^{(t-1)}$

$$I'_{RF} \leftarrow \mathrm{Repeat\_interleaved}(F^{(t)}), \qquad I_{SS} \leftarrow I'_{RF} \cdot S_m, \qquad G_k \leftarrow \mathrm{FFT}(I_{SS})$$

$I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$0

where $I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$1 is a learnable step size, $I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$2 is the subsampling mask, and $I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$3 is the acquired undersampled k-space (Hamedani et al., 28 Aug 2025). This update reflects the unrolled optimization view: a learned reconstruction term $I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$4 is combined with a measurement-consistency correction that pushes the estimate back toward the observed k-space data (Hamedani et al., 28 Aug 2025).

The paper states that the reconstructors are APUnet modules and that prompt learning, inspired by PromptMR, is incorporated to support adaptive reconstruction across different contrasts, trajectories, and anatomical views within one model (Hamedani et al., 28 Aug 2025). However, the implementation specifics of the prompt design are not fully detailed (Hamedani et al., 28 Aug 2025). Likewise, the paper does not provide exact numbers of convolution layers per APUnet, kernel sizes, feature channel counts, normalization layers, activation functions, or discriminator depth (Hamedani et al., 28 Aug 2025). This suggests that the paper’s main architectural contribution lies more in the residual unrolled design and its integration with domain-generalization losses than in low-level module novelty.

3. Objective function and domain-generalization mechanisms

GENRE-CMR defines a composite objective consisting of fidelity, Edge-Aware Region loss, and Statistical Distribution Alignment loss (Hamedani et al., 28 Aug 2025):

$I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$5

In practice, adversarial loss is also used through GAN training, so the generator objective is described as

$I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$6

(Hamedani et al., 28 Aug 2025).

The fidelity term combines a physical k-space consistency term and an image-domain SSIM loss (Hamedani et al., 28 Aug 2025). For each stage $I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$7,

$I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$8

$I^{(ACS)} \leftarrow \mathrm{FFT^{-1}}(k^{(ACS)})$9

$$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$0

where $$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$1 denotes coil combination, $$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$2 and $$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$3 denote magnitude and phase, and $$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$4 is ground-truth k-space (Hamedani et al., 28 Aug 2025).

The Edge-Aware Region loss is a central contribution. It is motivated by the tendency of deep reconstructions to appear smooth while blurring diagnostically important boundaries such as myocardial borders and small vessels (Hamedani et al., 28 Aug 2025). Given ground-truth image $$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$5, Sobel gradients are computed:

$$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$6

and the edge magnitude map is

$$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$7

(Hamedani et al., 28 Aug 2025). To enlarge the neighborhood around edges, the map is smoothed with a $$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$8 averaging kernel $$S_m \leftarrow \mathrm{APUnet}^{(SME)}(I^{(ACS)}), \qquad S_m' \leftarrow \mathrm{Conjugate\_Symmetry}(S_m)$$9:

$F^{(-1)} \leftarrow 0$0

A binary mask is then defined with threshold $F^{(-1)} \leftarrow 0$1:

$F^{(-1)} \leftarrow 0$2

The mask is applied to reconstructed and ground-truth images:

$F^{(-1)} \leftarrow 0$3

and the loss is

$F^{(-1)} \leftarrow 0$4

(Hamedani et al., 28 Aug 2025). The intended effect is to penalize structural degradation near edges rather than treating all pixels equally.

The Statistical Distribution Alignment loss is the most explicit generalization mechanism (Hamedani et al., 28 Aug 2025). The training data are assumed to come from five domains,

$F^{(-1)} \leftarrow 0$5

and groups of five consecutive samples are organized so that each group includes one sample from each distribution (Hamedani et al., 28 Aug 2025). For each sample $F^{(-1)} \leftarrow 0$6, feature vectors are extracted from each reconstructor $F^{(-1)} \leftarrow 0$7:

$F^{(-1)} \leftarrow 0$8

For each reconstructor $F^{(-1)} \leftarrow 0$9, the domain-specific feature set is

$t \in \{0,\dots,T-1\}$0

Assuming each domain’s features at layer $t \in \{0,\dots,T-1\}$1 follow a Gaussian $t \in \{0,\dots,T-1\}$2, the SDA loss at layer $t \in \{0,\dots,T-1\}$3 is

$t \in \{0,\dots,T-1\}$4

and the total SDA term is

$t \in \{0,\dots,T-1\}$5

(Hamedani et al., 28 Aug 2025). In implementation, the aligned feature is obtained by global average pooling,

$t \in \{0,\dots,T-1\}$6

and a feature bank with a sliding-window mechanism is used so that after the initial five samples, each new input is compared against the four most recent samples (Hamedani et al., 28 Aug 2025).

The discriminator is trained with binary cross-entropy losses. Real and fake discriminator inputs are formed by concatenating either the ground-truth image or the reconstructed image with the zero-filled or undersampled reference image (Hamedani et al., 28 Aug 2025):

$t \in \{0,\dots,T-1\}$7

$t \in \{0,\dots,T-1\}$8

$t \in \{0,\dots,T-1\}$9

$T=16$0

(Hamedani et al., 28 Aug 2025).

Finally, the weights $T=16$1 are not fixed constants. GENRE-CMR uses Coefficient of Variation weighting,

$T=16$2

with dynamically updated weights rather than reported fixed values (Hamedani et al., 28 Aug 2025).

4. Generalization strategy in relation to broader CMR research

GENRE-CMR’s claim to generalizability rests on three interacting mechanisms: residual feature propagation across cascades, EAR loss for structure-preserving robustness, and SDA loss for domain-invariant latent statistics (Hamedani et al., 28 Aug 2025). Residual propagation preserves and reuses early-stage features, EAR anchors supervision around comparatively stable structural cues such as edges, and SDA discourages learning domain-specific internal representations tied to vendor, trajectory, contrast, or protocol (Hamedani et al., 28 Aug 2025).

This design is consistent with several broader patterns in CMR research. Earlier generative restoration work used adversarial realism priors and anatomy-relevant auxiliary losses to sharpen clinically important boundaries in post-reconstruction CMR correction (Zhang et al., 2019). Large-scale cine CMR analysis pipelines have shown that multi-center, multi-vendor robustness depends not only on network design but also on harmonization, heterogeneity-aware training, and explicit quality control (Mariscal-Harana et al., 2022). Multimodal representation-learning studies have likewise shown that clinically informative CMR embeddings can benefit from alignment losses across sequences, especially when one modality is used to enrich another during training (Yuan et al., 23 Aug 2025). Registration-aware multi-sequence CMR segmentation further suggests that common-space alignment and anatomy-guided decoding are useful when domain or sequence mismatch degrades direct feature fusion (Ding et al., 2023).

These related results do not define GENRE-CMR itself, but they place it within a technical lineage in which residual generators, structure-aware supervision, adversarial learning, and alignment objectives are increasingly used to manage heterogeneity in CMR data (Zhang et al., 2019, Mariscal-Harana et al., 2022, Yuan et al., 23 Aug 2025, Ding et al., 2023). A plausible implication is that GENRE-CMR’s distinct contribution is to combine these tendencies in the specific setting of accelerated reconstruction under multi-domain shift.

5. Experimental protocol and empirical results

Experiments are conducted on the CMRxRecon 2025 dataset, described as a large-scale, multi-center, multi-vendor benchmark for robust cardiac MRI reconstruction (Hamedani et al., 28 Aug 2025). The dataset contains over 600 subjects, includes institutions and scanner vendors GE, Philips, Siemens, and UIH, includes healthy volunteers and patients with cardiomyopathies, myocardial infarction, arrhythmias, and pediatric cases, and spans cine, T1/T2 mapping, LGE, and perfusion (Hamedani et al., 28 Aug 2025). Sampling trajectories include Cartesian, radial, and Gaussian; field strengths include 1.5T, 3T, and 5T; and the dataset provides fully sampled and undersampled k-space, masks, and ground-truth reconstructions (Hamedani et al., 28 Aug 2025). The paper distinguishes training distributions and unseen distributions but does not provide exact train/validation/test counts or define exactly which domains are held out as unseen (Hamedani et al., 28 Aug 2025).

Implementation details reported for GENRE-CMR are: AdamW optimizer, learning rate 0.002, weight decay 0.1, gradient clipping 0.1, step scheduler with step size 11 and gamma 0.1, 16 reconstructor modules, 16 auto-calibration lines, adjacent k-space length 5, acceleration factors 8, 16, and 24, sampling trajectories k-t uniform, k-t Gaussian, and k-t radial, curriculum learning starting from lower acceleration factors, training duration 20 epochs, and batch size 1 (Hamedani et al., 28 Aug 2025).

GENRE-CMR is compared against PromptMR, SR-GAN, and PromptMR+ using SSIM, PSNR, and NMSE (Hamedani et al., 28 Aug 2025). The main results are as follows:

Setting	Method	SSIM	PSNR	NMSE
Training distributions	PromptMR	0.9685	41.80	0.0129
Training distributions	SR-GAN	0.9702	42.05	0.0120
Training distributions	PromptMR+	0.9728	42.40	0.0115
Training distributions	GENRE-CMR	0.9743	42.64	0.0111
Unseen distributions	PromptMR	0.9450	37.85	0.0198
Unseen distributions	SR-GAN	0.9473	38.01	0.0191
Unseen distributions	PromptMR+	0.9498	38.22	0.0187
Unseen distributions	GENRE-CMR	0.9552	38.90	0.0160

On unseen distributions, the paper highlights that GENRE-CMR improves over PromptMR+ by about $T=16$3 SSIM, about $T=16$4 dB PSNR, and reduces NMSE from 0.0187 to 0.0160 (Hamedani et al., 28 Aug 2025). Qualitatively, the model is reported to produce sharper anatomical boundaries, reduced aliasing artifacts, better preservation of myocardial borders and small vessels, and less blur under different trajectories at acceleration factor 16 (Hamedani et al., 28 Aug 2025).

The ablation study on unseen distributions isolates the contributions of the proposed components (Hamedani et al., 28 Aug 2025):

Variant	SSIM	PSNR	NMSE
Baseline	0.9473	38.01	0.0191
Without SDA Loss	0.9500	38.25	0.0183
Without EAR Loss	0.9515	38.43	0.0178
Without Residual Connections	0.9523	38.21	0.0171
Proposed Method	0.9552	38.90	0.0160

These results support the paper’s interpretation that the three mechanisms are complementary (Hamedani et al., 28 Aug 2025). Removing residual connections lowers unseen-distribution performance from SSIM 0.9552 / PSNR 38.90 / NMSE 0.0160 to SSIM 0.9523 / PSNR 38.21 / NMSE 0.0171; removing EAR lowers it to SSIM 0.9515 / PSNR 38.43 / NMSE 0.0178; and removing SDA lowers it to SSIM 0.9500 / PSNR 38.25 / NMSE 0.0183 (Hamedani et al., 28 Aug 2025). The largest NMSE penalty appears when SDA is removed, which the authors take as support for its domain-generalization role (Hamedani et al., 28 Aug 2025).

6. Significance, limitations, and interpretation

GENRE-CMR is presented as a step toward clinically adaptable deployment across heterogeneous acquisition protocols (Hamedani et al., 28 Aug 2025). Its intended practical appeal is that one model should work across multiple vendors, trajectories, contrasts, field strengths, and anatomical views or domains (Hamedani et al., 28 Aug 2025). The paper does not provide explicit inference runtime numbers, but it states that the framework is computationally demanding during training and requires powerful GPUs for optimal performance (Hamedani et al., 28 Aug 2025).

Several limitations are explicitly acknowledged or directly implied. The paper notes high computational cost during training, need for powerful GPUs, and lack of direct clinical validation so far (Hamedani et al., 28 Aug 2025). It also proposes future work including more efficient architectures or training strategies, radiologist-based clinical evaluation for diagnostic assessment, and movement toward real clinical integration (Hamedani et al., 28 Aug 2025). Other limitations are implicit: some architectural details are under-specified, the exact unseen-domain split is not fully described, and the contributions of adversarial training and prompt learning are not isolated in ablation (Hamedani et al., 28 Aug 2025).

From an encyclopedic perspective, GENRE-CMR is best understood as a multi-domain, supervised, GAN-based, residual unrolled CMR reconstruction framework whose main technical innovations for generalization are residual feature propagation in the unrolled generator, edge-aware structural supervision via EAR, and latent cross-domain alignment via SDA (Hamedani et al., 28 Aug 2025). Its strongest evidence is empirical performance on the CMRxRecon 2025 benchmark, particularly the unseen-distribution result of 0.9552 SSIM and 38.90 dB PSNR (Hamedani et al., 28 Aug 2025). This suggests that robust CMR reconstruction across heterogeneous clinical domains may benefit from combining physics-aware unrolling with explicit structure preservation and feature-distribution alignment, rather than treating generalization as a secondary effect of reconstruction fidelity alone.