7T-Restormer: 7T T1 Map Synthesis from 1.5/3T MRI

• The paper introduces a supervised paired image-to-image translation system that predicts 7T MP2RAGE-derived T1 maps from lower-field MRI, emphasizing quantitative relaxometry.
• It utilizes a Restormer-based encoder-decoder architecture with efficient MDTA and GDFN modules, achieving lower NMSE and higher PSNR compared to ResShift and ResViT.
• The method offers a computational surrogate to direct 7T acquisition for improved cortical and lesion visualization in multiple sclerosis, despite limitations in 2D context and sample diversity.

Searching arXiv for the specified paper and closely related work to ground the encyclopedia entry. 7T-Restormer is a lightweight transformer-based image synthesis model for predicting synthetic 7T quantitative T1 maps from routine 1.5T or 3T T1-weighted MRI. It is presented as a supervised paired image-to-image translation system in which a single-channel 2D lower-field T1-weighted slice is mapped to a single-channel 2D 7T MP2RAGE-derived T1 map, with the aim of making some benefits of ultra-high-field MRI available without requiring direct 7T acquisition (Eidex et al., 11 Jul 2025). The method is positioned specifically around quantitative relaxometry rather than generic “7T-like” appearance synthesis, and its reported application domain is multiple sclerosis.

1. Definition and task formulation

7T-Restormer addresses ultra-high-field MRI synthesis by learning a mapping from lower-field structural MRI to a quantitative 7T target. The paper formulates the task as supervised paired image-to-image translation: given an input axial slice

$x \in \mathbb{R}^{H \times W},$

the network learns a mapping

$F_\theta: x \mapsto \hat{y}$

to predict a paired 7T quantitative T1 map $y$, minimizing voxelwise $L_1$ reconstruction loss,

$\| y - \hat{y} \|_1 .$

The input modality is a single-channel 2D T1-weighted MRI slice from either 1.5T or 3T, and the output is a single-channel 2D synthesized 7T T1 map derived from 7T MP2RAGE (Eidex et al., 11 Jul 2025).

The target is therefore not a 7T MPRAGE image or a generic high-field anatomical contrast. It is specifically a 7T MP2RAGE-derived T1 relaxometry map. This distinction is central because the paper frames the method as quantitative map synthesis rather than conventional style transfer. The authors also report that pure $L_1$ loss gave lower NMSE than combining adversarial and $L_1$ losses, and that it reduced hallucinated textures and improved training stability, although at the cost of some smoothing in uncertain regions (Eidex et al., 11 Jul 2025).

The work is presented in the context of 7T MRI’s improved resolution and contrast relative to standard clinical field strengths, while acknowledging the practical constraints of true 7T acquisition, including scanner cost, limited availability, technical complexity, and susceptibility to artifacts. A plausible implication is that the model is intended less as a replacement for 7T physics than as a computational surrogate for some of its image-derived quantitative outputs.

2. Clinical motivation and scope of application

The clinical motivation is tied to the use of 7T MP2RAGE for cortical characterization and quantitative analysis. The paper states that at 7T, MP2RAGE reduces sensitivity to $B_1^+$ and $B_0$ inhomogeneity, and it places the work in neurologic disease, especially multiple sclerosis, where improved lesion depiction, cortical parcellation, and tissue characterization may improve diagnosis and monitoring (Eidex et al., 11 Jul 2025).

The retrospective dataset consisted of paired lower-field and 7T scans from patients with confirmed multiple sclerosis. The paper reports 108 patients with paired 3T and 7T scans and 35 patients with paired 1.5T and 7T scans, but also states that 2 patients were removed from the test dataset due to severe artifacts, leaving a final cohort of 141 paired cases. This nominal arithmetic inconsistency is present in the paper itself. The random patient-level split used for experiments was 105 training cases, 19 validation cases, and 17 test cases, with corresponding slice counts of 19,204, 3,476, and 3,145 (Eidex et al., 11 Jul 2025).

The lower-field scans came from various scanners and sequences, while the 7T target maps were acquired on a Siemens MAGNETOM Terra 7T scanner using an 8-channel transmit / 32-channel receive head coil and a 3D Low Angle Minimizing Artifacts MP2RAGE sequence. Reported imaging parameters for the 7T target included $TR = 4.5$ s, $F_\theta: x \mapsto \hat{y}$0 ms, $F_\theta: x \mapsto \hat{y}$1 s, matrix size $F_\theta: x \mapsto \hat{y}$2, and resolution $F_\theta: x \mapsto \hat{y}$3 mm$F_\theta: x \mapsto \hat{y}$4. The text also reports $F_\theta: x \mapsto \hat{y}$5, while noting that the manuscript appears ambiguous at that point, and gives the field of view as $F_\theta: x \mapsto \hat{y}$6 cm$F_\theta: x \mapsto \hat{y}$7 as written (Eidex et al., 11 Jul 2025).

This disease-specific and institution-specific framing constrains the scope of the results. The paper explicitly acknowledges that performance on other pathologies such as glioblastoma, or on broader neurologic populations, remains unknown (Eidex et al., 11 Jul 2025).

3. Data preparation and experimental protocol

Preprocessing consisted of rigid registration of lower-field MRI into the 7T image space, resampling to $F_\theta: x \mapsto \hat{y}$8 mm isotropic resolution, skull stripping using FSL BET, extraction of 2D axial slices, center cropping to $F_\theta: x \mapsto \hat{y}$9, intensity normalization to $y$0, and application of a Lipari colormap to the 7T images following Fuderer et al. The model therefore operates slice-wise in 2D rather than on 3D volumes (Eidex et al., 11 Jul 2025).

The train/validation/test partitions used mixed 1.5T and 3T inputs. The training cohort contained 25 cases with 1.5T input and 80 with 3T input; the validation cohort 5 with 1.5T and 14 with 3T; and the test cohort 5 with 1.5T and 14 with 3T. The abstract states a total of 32,128 slices, whereas the split totals sum to 25,825. This inconsistency is explicitly present in the paper, and the methods-section split counts are the operationally relevant numbers because they define the actual experiments (Eidex et al., 11 Jul 2025).

Training used AdamW with learning rate $y$1, $y$2, $y$3, weight decay $y$4, batch size 8, full precision, and 50 epochs on a single NVIDIA A6000 Ada with 48 GB. Data augmentation consisted of random flipping of axial slices “across the axial plane.” The paper does not explicitly describe a learning-rate schedule, model selection rule, or early stopping protocol (Eidex et al., 11 Jul 2025).

Evaluation used NMSE, PSNR, and SSIM. Intensities were linearly rescaled to $y$5 before metric calculation, so the dynamic range was

$y$6

For SSIM, the constants were

$y$7

The paper states that statistical significance was assessed with a two-tailed Welch’s $y$8-test, with significance threshold $y$9, and that all comparisons of 7T-Restormer versus baselines were significant with $L_1$0 (Eidex et al., 11 Jul 2025).

4. Architecture and computational design

7T-Restormer is an encoder-decoder built on the Restormer backbone. It takes a single-channel $L_1$1 axial T1-weighted slice and outputs a synthesized single-channel 7T T1 map of the same size (Eidex et al., 11 Jul 2025).

The structure begins with a $L_1$2 overlapping convolution that maps the input to 48 channels. Downsampling is performed with pixel-unshuffle, giving a channel progression of $L_1$3, with two downsampling stages corresponding to spatial resolutions $L_1$4, $L_1$5, and $L_1$6. The bottleneck contains an additional global MDTA block. Upsampling is performed by pixel-shuffle, and decoder features are concatenated with corresponding encoder features through skip connections. The output head consists of a final convolution followed by tanh activation, producing output normalized to $L_1$7 (Eidex et al., 11 Jul 2025).

Each Restormer block contains two residualized submodules in sequence:

1. MDTA, or Multi-Dconv Head Transposed Attention
2. GDFN, or Gated-Dconv Feed-Forward Network

A faithful Restormer-style block is reconstructed in the paper summary as

$L_1$8

$L_1$9

The manuscript itself states only “normalization layer,” without specifying the exact LayerNorm variant (Eidex et al., 11 Jul 2025).

MDTA is described as the core efficiency mechanism. The normalized feature map is projected to $\| y - \hat{y} \|_1 .$0, $\| y - \hat{y} \|_1 .$1, and $\| y - \hat{y} \|_1 .$2 using a $\| y - \hat{y} \|_1 .$3 pointwise convolution followed by a $\| y - \hat{y} \|_1 .$4 depthwise convolution. Attention is computed across channels rather than with full spatial self-attention, and the paper describes this as reducing complexity from

$\| y - \hat{y} \|_1 .$5

to

$\| y - \hat{y} \|_1 .$6

The practical rationale is that global anatomical dependencies are modeled without the quadratic spatial cost of standard self-attention (Eidex et al., 11 Jul 2025).

GDFN follows MDTA and uses two parallel depthwise-separable pathways, one with GELU and one linear, whose outputs are combined through elementwise multiplication as a gating mechanism, followed by a $\| y - \hat{y} \|_1 .$7 convolution and residual addition. This feed-forward design is intended to preserve local spatial structure efficiently (Eidex et al., 11 Jul 2025).

The model’s reported efficiency is one of its main distinguishing features. 7T-Restormer has 10.5 million trainable parameters, compared with 56.7 million for ResViT and 70.4 million for ResShift. Inference time is reported as 0.27 s to generate one $\| y - \hat{y} \|_1 .$8 axial slice on an NVIDIA A6000 GPU (Eidex et al., 11 Jul 2025).

5. Empirical performance and ablation results

The main baselines were ResViT and ResShift. For 1.5T inputs, 7T-Restormer achieved NMSE $\| y - \hat{y} \|_1 .$9, PSNR $L_1$0 dB, and SSIM $L_1$1. For 3T inputs, it achieved NMSE $L_1$2, PSNR $L_1$3 dB, and SSIM $L_1$4 (Eidex et al., 11 Jul 2025).

For comparison, the reported baseline values were:

Input	Method	NMSE / PSNR / SSIM
1.5T	ResShift	$L_1$5, $L_1$6 dB, $L_1$7
1.5T	ResViT	$L_1$8, $L_1$9 dB, $L_1$0
3T	ResShift	$L_1$1, $L_1$2 dB, $L_1$3
3T	ResViT	$L_1$4, $L_1$5 dB, $L_1$6

The paper emphasizes that 7T-Restormer achieved the best NMSE and PSNR across both input field strengths, while ResShift slightly exceeded it in SSIM. The relative NMSE reductions reported were 64% versus ResShift and 41% versus ResViT at 1.5T, and 55% versus ResShift and 36% versus ResViT at 3T (Eidex et al., 11 Jul 2025).

Qualitatively, the reported findings are that 7T-Restormer better preserved cortical-subcortical contrast, sulcal pattern delineation, ventricular CSF intensity, conformity in inferior brain, cerebellar, or brainstem regions, and reduced spurious ring-like hallucinations seen in ResViT in some difficult areas. ResShift is described as tending to oversharpen edges and exaggerate gray-white boundaries, whereas ResViT is described as tending to oversmooth structures such as deep gray nuclei and occasionally inject noise or artifacts. The paper also notes that all methods can sometimes copy details from the input T1-weighted image that are not actually present in the 7T T1 target, and all methods may oversmooth or fail on very complex structures (Eidex et al., 11 Jul 2025).

Ablation experiments examined mixed-field training versus single-field training. On 1.5T test slices, mixed 1.5T+3T training yielded NMSE $L_1$7, outperforming both 1.5T-only and 3T-only training, each of which produced NMSE $L_1$8. On 3T test slices, mixed training matched 3T-only training on NMSE and slightly exceeded 1.5T-only training. The paper concludes that mixed-field training was superior to single-field strategies overall, and suggests that contrast and noise differences across field strengths were complementary rather than harmful (Eidex et al., 11 Jul 2025).

This suggests that the model’s generalization mechanism is not purely field-specific. A plausible implication is that mixed-field supervision regularizes the mapping from conventional T1-weighted anatomy to 7T quantitative T1 structure more effectively than restricting the network to a single input field strength.

6. Interpretation, misconceptions, related work, and limitations

Several interpretive points are explicitly supported by the paper. First, the method synthesizes 7T MP2RAGE-derived T1 maps from lower-field T1-weighted MRI; it does not acquire or reconstruct physically measured 7T T1 values. The paper therefore states that the output is a synthetic approximation of a quantitative 7T T1 map, not a direct measurement in the physical sense (Eidex et al., 11 Jul 2025).

Second, the model is 2D and slice-based. It ignores through-plane context, which the paper identifies as a limitation relative to possible 3D architectures. The authors argue that, for the dataset size available, a 2D approach may be more practical because 3D training would drastically reduce the effective sample count (Eidex et al., 11 Jul 2025).

Third, the evidence is strongest for image-level fidelity rather than clinical utility. The paper supports that 7T-Restormer yields lower NMSE and higher PSNR than ResViT and ResShift, with far fewer parameters, on the reported test set. By contrast, improvement in downstream diagnosis, treatment planning, segmentation, or parcellation is more suggestive than proven. Generalization across institutions and vendors is also not established, even though the lower-field scans were heterogeneous, because the study appears single-institution and the 7T target acquisition used a single vendor/platform (Eidex et al., 11 Jul 2025).

The work also sits within a broader literature on 7T MRI synthesis and restoration. A related but distinct model is the hybrid Vision CNN-Transformer for 3T-to-7T ADC synthesis, which used 3T ADC and 3T T1-weighted MRI to synthesize 7T ADC maps and reported hold-out results of MSE $L_1$9, PSNR $B_1^+$0 dB, and SSIM $B_1^+$1 (Eidex et al., 2023). Another later approach, FS-RWKV, framed paired 3T-to-7T MRI translation as a frequency-spatial RWKV problem and reported superior performance over CNN-, GAN-, Transformer-, and RWKV-based baselines across T1w and T2w synthesis on UNC and BNU datasets (Lei et al., 10 Oct 2025). These comparisons are architectural rather than directly experimental with respect to 7T-Restormer, because the datasets, targets, and modalities differ.

A common misconception is to conflate any 7T synthesis or restoration pipeline with a Restormer. The 7T $B_1^+$2-separation work centered on R2PRIMEnet7T, for example, used a patch-based 3D U-net to predict a 3T-range $B_1^+$3 surrogate from 7T $B_1^+$4, and explicitly did not contain transformer blocks or Restormer-style attention (Kim et al., 2024). By contrast, 7T-Restormer is literally a Restormer-based encoder-decoder and is directly tied to transformer-style transposed attention (Eidex et al., 11 Jul 2025).

The paper’s acknowledged limitations are that the data are entirely from patients with multiple sclerosis, that supervised training requires paired lower-field and 7T data, that the model is 2D, that the 7T reference comes from a single vendor platform, and that the output may miss subtle pathology, oversmooth complex structures, or import structures from the source T1-weighted image that are not perfectly matched to the 7T target (Eidex et al., 11 Jul 2025). Future directions proposed in the paper include pretraining on large-scale brain MRI corpora, foundation-model style transfer learning, 3D architectures, synthesis of other 7T sequences, larger and more diverse patient cohorts, and multisite validation (Eidex et al., 11 Jul 2025).

In technical terms, 7T-Restormer is best understood as a Restormer-derived quantitative map synthesizer for 1.5T/3T-to-7T translation, optimized for low parameter count and stable supervised training. Its central claim is not universal superiority across every image metric, but lower reconstruction error and higher PSNR than the tested state-of-the-art comparators, with a substantially smaller model footprint (Eidex et al., 11 Jul 2025).