HierAdaptMR: Hierarchical Cardiac MRI Reconstruction

• The paper demonstrates that HierAdaptMR integrates a variational unrolling backbone with protocol and center-level adapters, yielding notable SSIM and PSNR improvements.
• HierAdaptMR is a deep-learning framework that decomposes domain shifts into protocol-specific, center-specific, and universal adaptations to enable efficient cross-center generalization.
• Ablation studies show that the hierarchical adapter structure enhances artifact suppression and image sharpness, achieving robust performance across diverse scanner configurations.

HierAdaptMR most directly denotes a hierarchical feature adaptation framework for cross-center cardiac MRI reconstruction. In that formulation, a variational unrolling backbone is combined with protocol-level adapters, center-level adapters, and a universal adapter for unseen centers, with the stated goal of handling multi-level domain variations induced by heterogeneous scanners, clinical centers, and imaging protocols (Xu et al., 18 Aug 2025). In adjacent usage, the same label has also been associated with hierarchical recurrent adapters for efficient multi-task adaptation of large speech models, and with broader interpretive notions of adaptive model hierarchies or hierarchical adaptive multi-retrieval; these usages are related by architectural motifs rather than by a single shared formalism (Munkhdalai et al., 2024).

1. Terminology and scope

Within the supplied literature, the name HierAdaptMR is used most explicitly as the title of "HierAdaptMR: Cross-Center Cardiac MRI Reconstruction with Hierarchical Feature Adapters," where it refers to a deep-learning framework for accelerated cardiac MRI reconstruction across multiple clinical centers with heterogeneous scanners and imaging protocols (Xu et al., 18 Aug 2025). Its defining components are a variational unrolling reconstruction backbone, protocol-specific and center-specific feature adapters, and a universal adapter intended for centers not seen during training.

The term is not uniformly canonical across all domains. In the speech adaptation literature, the query term is described as referring to the Hierarchical Recurrent Adapter (HRA) or to a very close implementation variant, rather than to a separately named method. In that setting, the core idea is a single shared recurrent controller plus lightweight task-specific heads for multi-task automatic speech recognition adaptation (Munkhdalai et al., 2024). In other words, the label can denote either a specific MRI reconstruction framework or, in looser usage, a family resemblance built around hierarchical, parameter-efficient adaptation.

This ambiguity matters because superficially similar names conceal substantially different technical objects. The cardiac MRI version is an image reconstruction system operating on undersampled k-space and post-reconstruction image features. The speech version is an adapter architecture for a large pre-trained ASR backbone. This suggests that “HierAdaptMR” is best treated as a term whose meaning depends on context, with the MRI framework as the most direct named instance in the supplied corpus.

2. Cross-center cardiac MRI problem formulation

In the cardiac MRI setting, HierAdaptMR addresses cross-center domain shift under accelerated acquisition. The stated sources of variation include center-level differences in field strength, vendor, coil configurations, receiver channels, and acquisition implementations; protocol-level differences across cine, LGE, T1/T2 mapping, perfusion, T1rho, Flow2d, BlackBlood, T1w, and T2w; and acquisition-level differences such as uniform, radial, and k-t Gaussian undersampling with acceleration factors from $8\times$ to $24\times$ (Xu et al., 18 Aug 2025).

The framework conceptually separates three kinds of domains. Protocol-level domains are defined by imaging sequence or modality. Center-level domains are defined by scanner and clinical center. Unseen centers are centers absent from training, for which protocol metadata remains available at test time. The stated deployment problem is therefore not only reconstruction from undersampled data, but reconstruction under a structured domain shift in which protocol identity is known while center identity may be novel.

The backbone reconstruction is framed as a variational unrolling process in k-space. Let $\mathbf{y}$ denote measured undersampled k-space, $\mathbf{k}^{(t)}$ the current k-space estimate at unrolling step $t$, $\mathcal{A}$ the forward encoding operator, $\mathcal{A}^H$ its adjoint, $\mathcal{M}$ the sampling mask, $\mathcal{R}_{\theta_t}$ a learnable regularization network, and $\lambda_t$ learnable step-size weights. The update rule is

$24\times$0

This is paired with an optimization view consisting of a data-fidelity term and a learned regularization term, and is implemented with 12 cascades (Xu et al., 18 Aug 2025).

The same paper states that the backbone is an optimized combination of HyperCMR and PromptMR-plus. PromptMR-plus contributes a historical feature memory across cascades, with enhanced features at cascade stage $24\times$1 defined by

$24\times$2

where $24\times$3 are learned attention weights and $24\times$4 is a memory aggregation function (Xu et al., 18 Aug 2025). The reconstructed image is then obtained from the final k-space estimate using inverse Fourier transformation and RSS coil combination.

3. Hierarchical adapter architecture

The hierarchical component of HierAdaptMR is implemented as post-reconstruction feature-level adaptation on the reconstructed image $24\times$5. Three adapter classes are defined: a Protocol-Level Adapter $24\times$6, a Center-Level Adapter $24\times$7, and a Universal Adapter $24\times$8 (Xu et al., 18 Aug 2025). During training on known centers, the active transformation is composed as

$24\times$9

and the final reconstruction is

$\mathbf{y}$0

Each adapter is a lightweight residual module,

$\mathbf{y}$1

where $\mathbf{y}$2 is instance normalization, $\mathbf{y}$3 denotes convolution, $\mathbf{y}$4 is batch norm plus ReLU, $\mathbf{y}$5 are convolutional kernels, and $\mathbf{y}$6 is a scalar scaling factor initialized small, with $\mathbf{y}$7 stated in the implementation details (Xu et al., 18 Aug 2025). The channel dimensions are $\mathbf{y}$8, preserving spatial resolution while progressively compressing feature dimensionality.

The protocol-level adapters are sequence-specific modules selected deterministically from metadata. The center-level adapters are center-specific modules for centers seen during training. The universal adapter is a center-invariant transformation intended for zero-shot use on unseen centers. The supplied details state that protocol adapters contribute approximately $\mathbf{y}$9K parameters per protocol, center adapters approximately $\mathbf{k}^{(t)}$0K parameters per center, and the total adapter parameters approximately $\mathbf{k}^{(t)}$1M, equal to $\mathbf{k}^{(t)}$2 of the backbone, yielding approximately $\mathbf{k}^{(t)}$3 parameter savings relative to full fine-tuning (Xu et al., 18 Aug 2025).

Architecturally, the method treats domain shift as multi-level rather than monolithic. Protocol-level modules target contrast behavior, temporal dynamics, and modality-specific texture. Center-level modules target scanner-dependent intensity and artifact biases. The universal adapter absorbs what is described as center-invariant adaptation. This decomposition is central to the method’s claim that cross-center generalization requires more than a single undifferentiated domain adapter.

4. Optimization, stochastic adapter selection, and generalization

The optimization strategy combines multi-scale structural similarity with a frequency-domain term and contrast-adaptive weighting. The multi-scale SSIM loss is defined over downsampling scales $\mathbf{k}^{(t)}$4, weighting smaller scales more heavily through a factor $\mathbf{k}^{(t)}$5, while the frequency-domain SSIM loss compares normalized magnitudes of Fourier transforms of reconstructed and target images (Xu et al., 18 Aug 2025). The total loss for contrast $\mathbf{k}^{(t)}$6 is described as

$\mathbf{k}^{(t)}$7

The supplied examples state that cine uses higher $\mathbf{k}^{(t)}$8, LGE uses higher $\mathbf{k}^{(t)}$9, and mapping sequences may reduce $t$0 (Xu et al., 18 Aug 2025).

Generalization to unseen centers is driven by stochastic selection between the center-specific and universal adapters. For known centers during training,

$t$1

The supplementary material is summarized as using $t$2, so known-center training applies the center-specific adapter $t$3 of the time and the universal adapter $t$4 of the time (Xu et al., 18 Aug 2025). At test time on unseen centers, the active composition becomes

$t$5

The paper also states that there is no explicit gating network. Adapter selection is purely metadata-driven, using protocol labels and center identifiers where available (Xu et al., 18 Aug 2025). This clarifies a potential misconception: the universal adapter is not a replacement for protocol metadata, nor does the framework infer protocol identity from the image alone. Instead, known protocol labels are assumed, while the unseen-center case is handled by replacing center-specific adaptation with the universal adapter.

Implementation details further specify PyTorch training with mixed precision on NVIDIA A100 80GB hardware, AdamW optimization with weight decay $t$6, cosine annealing to $t$7, early stopping with patience $t$8 epochs based on validation SSIM, gradient accumulation over 8 steps, and hierarchical stratified sampling on an approximately $t$9 subset that preserves Center $\mathcal{A}$0 Vendor $\mathcal{A}$1 Modality $\mathcal{A}$2 Patient structure (Xu et al., 18 Aug 2025).

5. Experimental results, ablations, and limitations

The experimental setting is the CMRxRecon2025 dataset, described as spanning $\mathcal{A}$3 clinical centers, $\mathcal{A}$4 scanners, and $\mathcal{A}$5 cardiac imaging modalities, with centers labeled C001–C008, vendors UIH, Siemens, GE, and Philips, field strengths $\mathcal{A}$6T–$\mathcal{A}$7T, and multiple anatomical views (Xu et al., 18 Aug 2025). The evaluation protocol includes center-based splitting, with one entire center held out as validation to simulate an unseen domain and approximately $\mathcal{A}$8 of the data assigned to validation.

Across centers and vendors, the summarized overall mean results report HyperCMR baseline at SSIM $\mathcal{A}$9 and PSNR $\mathcal{A}^H$0 dB, versus HierAdaptMR at SSIM $\mathcal{A}^H$1 and PSNR $\mathcal{A}^H$2 dB, corresponding to $\mathcal{A}^H$3 SSIM and $\mathcal{A}^H$4 PSNR (Xu et al., 18 Aug 2025). Per-center examples include C001 (UIH-3.0T), where SSIM improves from $\mathcal{A}^H$5 to $\mathcal{A}^H$6 and PSNR from $\mathcal{A}^H$7 to $\mathcal{A}^H$8, and GE-1.5T centers C005 and C008, with $\mathcal{A}^H$9–$\mathcal{M}$0 SSIM and approximately $\mathcal{M}$1–$\mathcal{M}$2 PSNR gains. Mean field-strength gains are stated as $\mathcal{M}$3 SSIM and $\mathcal{M}$4 PSNR at $\mathcal{M}$5T, and $\mathcal{M}$6 SSIM and $\mathcal{M}$7 PSNR at $\mathcal{M}$8T (Xu et al., 18 Aug 2025).

On challenge-style Task1 and Task2 comparisons, the supplied summary reports Ours (HierAdaptMR, 26.5% data) at SSIM $\mathcal{M}$9 and PSNR $\mathcal{R}_{\theta_t}$0, exceeding HyperCMR trained on all data, which is reported at SSIM $\mathcal{R}_{\theta_t}$1 and PSNR $\mathcal{R}_{\theta_t}$2 (Xu et al., 18 Aug 2025). Qualitative examples across 11 center configurations are summarized as showing suppression of radial streaks and uniform aliasing, preservation of ventricles, myocardial walls, and valvular regions, and recovery of high-frequency k-space components.

The ablation discussion emphasizes several points. HyperCMR plus adapters improves over backbone-only reconstruction. Removing or flattening protocol adapters is described as likely degrading contrast-specific performance. Removing center-specific adapters would force the universal adapter to absorb all center variability. Excluding the universal adapter would prevent zero-shot generalization to unseen centers. Removing the frequency-domain SSIM term is said to reduce sharpness and increase residual artifacts, particularly for high-acceleration radial and Gaussian patterns (Xu et al., 18 Aug 2025). These are reported as analyses rather than complete tabulated ablations in the main text.

The limitations section identifies dependence on correct protocol and center metadata, the need for diverse multi-center training data, the post-reconstruction placement of adapters rather than earlier k-space or feature-level intervention, and lack of explicit evaluation on extreme scanner configurations such as very high field $\mathcal{R}_{\theta_t}$3T or exotic sequences (Xu et al., 18 Aug 2025). The same section notes that centers with extremely poor SNR or highly non-standard protocols may remain challenging, and that poor tuning of the universal-adapter usage probability could lead to underfitting on unseen centers.

6. Related usages of the name in other research areas

A distinct but related usage appears in speech adaptation. In "Hierarchical Recurrent Adapters for Efficient Multi-Task Adaptation of Large Speech Models," the query term HierAdaptMR is described as referring to the Hierarchical Recurrent Adapter (HRA) or a very close implementation variant (Munkhdalai et al., 2024). HRA adapts a large pre-trained speech model by splitting adaptation into a single shared recurrent controller and lightweight task-specific heads, with recurrence across backbone depth:

$\mathcal{R}_{\theta_t}$4

and residual adaptation

$\mathcal{R}_{\theta_t}$5

In the reported ASR experiments, HRA is said to achieve $\mathcal{R}_{\theta_t}$6–$\mathcal{R}_{\theta_t}$7 parameter efficiency, and in single-task settings a $\mathcal{R}_{\theta_t}$8M-parameter linear-head HRA matches or slightly beats full fine-tuning of an approximately $\mathcal{R}_{\theta_t}$9B-parameter model, while FFN-head HRA reaches the best WER among adapters (Munkhdalai et al., 2024).

Two further papers use the name only interpretively. In "Adaptive Model Hierarchies for Multi-Query Scenarios," the supplied explanation treats “HierAdaptMR” as a possible shorthand for Hierarchical Adaptive Models for Multi-Requests, meaning a certified multi-fidelity hierarchy in which low-cost models are adaptively improved using high-fidelity evaluations (Kleikamp et al., 2024). In "Generalizable Agent Modeling for Agent Collaboration-Competition Adaptation with Multi-Retrieval and Dynamic Generation," the supplied explanation similarly treats the label as a possible hierarchical adaptive multi-retrieval view of MRDG within ACCA, emphasizing stable versus adaptive parameters, task-level versus interaction-level modeling, and episode-level versus timestep-level adaptation (Wang et al., 20 Jun 2025). These usages are explicitly interpretive rather than official nomenclature.

Taken together, these cross-domain appearances indicate that the name is associated less with one invariant architecture than with a recurring design pattern: hierarchical decomposition, parameter-efficient specialization, and explicit handling of structured variation across tasks, domains, agents, or centers. This suggests that “HierAdaptMR” functions as a context-sensitive label whose most precise meaning depends on whether the subject is cardiac MRI reconstruction, large-scale speech adaptation, adaptive model hierarchies, or multi-agent adaptation.