PaperlayerView exact revision on arXiv

Source-linked AI summary

DGCM-SSFA: Dual-Granularity Cross-Modulation and Sparse Strip Frequency Attention for Efficient 3T-to-7T MRI Synthesis

Zhichun Fan, Zhiqiang Chen, Lanling Liu, Lingjun Liu, Zhonghua Xie

arXiv:2610.29141v1cs.CV

TL;DR

Synthesizing 7T-like MRI from routine 3T scans could provide enhanced neuroimaging, but the task must recover fine anatomical detail across spatial scales and domains. DGCM-SSFA combines multi-level local refinement with bottleneck spatial-frequency attention in an 11M-parameter network. It achieves competitive structural fidelity, including 0.8431 SSIM on BNU T2w, while the 2D slice-wise formulation leaves inter-slice consistency unmodeled.

  • Problem

    3T-to-7T synthesis must recover fine anatomical details, including tissue-boundary edges, long-range dependencies, and subtle textures, while 7T access is restricted by cost and specialized requirements.

    Source: §I · p. 1

  • Method

    DGCM-SSFA is an 11M-parameter U-Net Transformer that combines multi-level DGCM local refinement with bottleneck SSFA spatial-frequency attention.

    Source: §I · p. 2

  • Results

    DGCM-SSFA attains 0.8431 SSIM on BNU T2w, approaching FS-RWKV (0.8624), with 11M parameters.

    Source: §V · p. 5

  • Takeaways & Limitations

    DGCM-SSFA offers a practical, parameter-efficient approach to 3T-to-7T synthesis with strong structural fidelity.

    Source: §V · pp. 5, 7

    The 2D slice-wise formulation does not exploit inter-slice anatomical consistency; a volumetric 3D extension could improve fidelity metrics.

    Source: §V · p. 6

Abstract

from arXiv · show

Synthesizing high-fidelity 7T MRI from routinely acquired 3T scans offers a practical pathway to enhanced neuroimaging without the prohibitive cost of ultra-high-field scanners. While existing deep learning approaches have achieved promising results, state-of-the-art methods typically rely on either large-scale architectures exceeding 45 million parameters or computationally intensive ensemble strategies, limiting their practicality in resource-constrained clinical environments. In this work, we propose a compact multi-domain attention framework, termed DGCM-SSFA, that achieves competitive synthesis quality with only 11M parameters. The proposed architecture integrates two complementary modules into a U-Net Transformer backbone. First, a Dual-Granularity Cross-Modulated Feature Modulator (DGCM) is embedded at multiple hierarchical levels to enhance local edge and texture details through Sobel-guided spatial attention and depthwise frequency-aware channel attention with cross-branch gating. Second, a Sparse Strip Frequency Attention (SSFA) module at the bottleneck captures global spatial context via strip-pooled attention with near-linear complexity, while simultaneously performing FFT-based spectral modulation in the frequency domain. These two modules operate synergistically with the backbone's native channel attention, forming a triple-granularity feature refinement pipeline. Extensive experiments on the UNC and BNU 3T-to-7T MRI datasets across T1w and T2w modalities demonstrate that DGCM-SSFA achieves strong structural fidelity, attaining an SSIM of 0.8431 on BNU T2w---approaching the best benchmark value of 0.8624---and 0.8100 on BNU T1w, while using substantially fewer parameters than Transformer-based alternatives such as ResViT (123.4M) and FS-RWKV (45.4M).

I. INTRODUCTION

7T MRI offers higher signal-to-noise ratio and spatial resolution than 3T, but its cost and specialized requirements restrict access. The paper proposes DGCM-SSFA to combine local, global, and frequency-domain attention for 3T-to-7T synthesis in a compact architecture.

I. INTRODUCTION

  • 7T MRI offers an order-of-magnitude gain in signal-to-noise ratio and spatial resolution over 3T, but costs and specialized requirements restrict access.The passage identifies hardware costs exceeding several million dollars, specialized site requirements, and stringent safety protocols.Source: §I · p. 1
  • 3T-to-7T synthesis must recover fine anatomical details while modeling tissue-boundary edges, long-range dependencies, and subtle textures.The task remains an open challenge.Source: §I · p. 1
  • CNNs, GANs, Transformers, and state space models each leave trade-offs in local detail, anatomical reliability, complexity, or frequency-selective refinement.CNNs have limited receptive fields; GANs risk spurious anatomy; quadratic Transformer attention can discard high-frequency details through downsampling; and state space models lack explicit frequency-selective refinement.Source: §I · p. 1
  • DGCM-SSFA addresses the gap by combining local edge-guided, global spatial, and frequency-domain attention in one architecture.The proposed framework uses a U-Net Transformer backbone with DGCM modules at multiple encoder levels and SSFA at the bottleneck.Source: §I · pp. 1–2
  • DGCM-SSFA uses 11M parameters and attains 0.8431 SSIM on BNU T2w, approaching the 0.8624 benchmark value.The contribution passage describes experiments on UNC and BNU across T1w and T2w modalities.Source: §I · p. 2

A. Deep Learning for Cross-Field MRI Synthesis

Prior cross-field MRI methods balance local reconstruction, perceptual quality, and global context with limits in receptive field, hallucination risk, or computational cost. DGCM-SSFA responds by combining edge-aware spatial, strip-pooled global, and frequency-domain attention through cross-branch gating.

A. Deep Learning for Cross-Field MRI Synthesis

  • CNN regression can oversmooth anatomical boundaries, adversarial methods risk hallucinating anatomy, and Transformer models such as ResViT use 123M parameters.State space models offer a more efficient alternative, while the cited review identifies a remaining gap for compact architectures.Source: A. Deep Learning for Cross-Field MRI… · p. 2

B. Attention Mechanisms in Image Restoration

  • Sobel filters add explicit edge-awareness to spatial attention, while frequency processing separates high-frequency edge and texture information from low-frequency illumination and contrast.Strip pooling reduces the attention token space from O(HW) to O(H + W) for global context aggregation.Source: B. Attention Mechanisms in Image… · p. 2

C. Multi-Domain and Multi-Granularity Feature Fusion

  • DGCM-SSFA integrates local edge-guided spatial, global strip-pooled spatial, and frequency-domain spectral attention through cross-branch and cross-domain gating.The passage contrasts this design with MRI synthesis frameworks that typically use at most two attention types.Source: C. Multi-Domain and Multi-Granularity… · p. 2

III. METHOD

DGCM-SSFA combines a U-Net-style encoder-decoder with hierarchical DGCM refinement and bottleneck SSFA attention. Skip connections preserve low-level structure, and a global residual connection produces the final 7T-like image.

A. Overall Pipeline

  • The encoder progressively downsamples feature maps through FFN modules and convolutional layers, with DGCM refining local edge and texture details at deeper semantic levels.The encoder and decoder use repeated Transformer blocks; the encoder downsamples via stride-2 convolution across three levels.Source: A. Overall Pipeline · pp. 2–3
  • At the bottleneck, SSFA captures global spatial context through strip-pooled attention and performs FFT-based spectral modulation.The decoder upsamples features and concatenates encoder skip connections to preserve low-level structural information.Source: A. Overall Pipeline · pp. 2–3
  • DGCM-SSFA uses a U-Net-style encoder-decoder, placing DGCM after encoder levels and refinement for local detail enhancement and a single SSFA at the bottleneck for global spatial-frequency attention.The framework uses a global residual connection to produce its output.Source: Figure 1 · p. 3
  • After decoder refinement, a fourth DGCM enhances detail before the network projects a residual image and adds it to the input.DGCM uses edge-aware spatial and frequency-aware channel branches; SSFA combines sparse strip attention and FFT-based spectral attention with cross-gated fusion.Source: A. Overall Pipeline · p. 3

B. DGCM: Dual-Granularity Cross-Modulated Feature Modulator

DGCM refines feature maps at four hierarchical positions through a local-global residual design. Its spatial branch uses activation, intensity, and Sobel-gradient information to produce spatial attention.

B. DGCM: Dual-Granularity Cross-Modulated Feature Modulator

  • DGCM is placed after the first three encoder levels and after refinement, producing a same-dimensional refined feature map at each position.The module follows a local-global residual design, with a lightweight residual block composed of two convolutional layers and ReLU activation.Source: B. DGCM: Dual-Granularity… · p. 3

1) Edge-Aware Spatial Branch:

  • The spatial branch forms a three-channel descriptor from channel-wise maximum, channel-wise mean, and the mean map's Sobel gradient magnitude.The Sobel kernels are fixed, non-trainable buffers and add no parameters for edge extraction.Source: 1) Edge-Aware Spatial Branch · pp. 3–4
  • A lightweight convolutional head maps the descriptor to a single-channel spatial attention map via sigmoid activation.The map modulates the feature map by element-wise multiplication with broadcasting.Source: 1) Edge-Aware Spatial Branch · p. 4

2) Frequency-Aware Channel Branch:

The frequency-aware channel branch combines pooled descriptors to produce channel attention, while cross-modulation fusion adaptively blends the spatial and channel outputs.

2) Frequency-Aware Channel Branch:

  • Three pooled descriptors, including a frequency-aware response, feed a two-layer MLP that produces channel attention.The frequency-aware descriptor comes from a depthwise convolution modulated by spatial attention and global average pooling of its absolute response.Source: 2) Frequency-Aware Channel Branch · p. 4

3) Cross-Modulation Fusion:

  • DGCM concatenates spatial and channel outputs, then uses separate fusion and gating pathways to form an integrated feature map and sigmoid mask.The fusion pathway uses a 1×1–3×3 convolution sequence.Source: 3) Cross-Modulation Fusion · p. 4
  • The gated fusion softly blends the integrated representation with the branches’ direct average, adapting selection to local feature characteristics.Source: 3) Cross-Modulation Fusion · p. 4

C. SSFA: Sparse Strip Frequency Attention

SSFA targets long-range anatomical dependencies at the bottleneck through parallel spatial and frequency pathways. Strip pooling reduces attention cost by shrinking the key-value token space.

C. SSFA: Sparse Strip Frequency Attention

1) Sparse Strip Spatial Attention:

  • Strip pooling compresses the key and value token space from HW to H + W, replacing dense attention’s O(H^2W^2) cost with O(HW(H + W)).The reduction is particularly substantial at the bottleneck, where spatial dimensions are smallest.Source: 1) Sparse Strip Spatial Attention · p. 4

2) Frequency Spectral Attention:

The spectral pathway applies channel attention to the magnitude of an rFFT spectrum while retaining its phase, then reconstructs the refined spectrum.

2) Frequency Spectral Attention:

3) Cross-Domain Gating:

SSFA couples its spatial and frequency pathways through mutual gating, then fuses the gated outputs and adds a global residual.

3) Cross-Domain Gating:

D. Training Objective

The study trains DGCM-SSFA with an L1 reconstruction objective and evaluates it on preprocessed UNC and BNU scans using PSNR, SSIM, and RMSE. Training uses a two-stage patch schedule and standard geometric augmentation.

D. Training Objective

  • L1 reconstruction loss compares synthesized 7T images with ground truth; the global residual formulation predicts the 3T-to-7T difference, which the authors say stabilizes training.L1 is selected for outlier robustness and sharper outputs in image restoration tasks.Source: D. Training Objective · p. 5

E. Implementation Details

  • Training runs for 30,000 iterations with AdamW, cosine annealing, and batch size 8 per GPU across two NVIDIA RTX 5090 GPUs.Patch size increases from 128 × 128 to 256 × 256 after 10,000 iterations; augmentation includes flipping, 90-degree rotation, and cropping.Source: E. Implementation Details · p. 5

A. Datasets and Preprocessing

  • UNC includes 10 participants and BNU 20; volumes are resampled to 1.0 mm, skull-stripped, rigidly registered, normalized to [0, 1], and sliced into 256 × 256 axial images.Both datasets contain paired 3T and 7T scans, with T1w and T2w modalities.Source: A. Datasets and Preprocessing · p. 5

B. Evaluation Metrics

  • PSNR, SSIM, and RMSE assess synthesis quality: PSNR and RMSE capture pixel-level fidelity, while SSIM reflects structural consistency.Higher PSNR and SSIM and lower RMSE indicate better reconstruction.Source: B. Evaluation Metrics · p. 5

C. Comparison with State-of-the-Art Methods

DGCM-SSFA delivers competitive structural similarity with 11M parameters, and visual comparisons report sharper boundaries and finer textures. Ablations on BNU T2w indicate contributions from both DGCM and SSFA, with largely orthogonal feature effects.

C. Comparison with State-of-the-Art Methods

  • 0.8431 SSIM on BNU T2w and 0.8100 on BNU T1w place DGCM-SSFA second to FS-RWKV, while its 11M parameters are fewer than several compared methods.FS-RWKV scores 0.8624 SSIM on BNU T2w and 0.8388 on BNU T1w; parameter counts include 45.40M for FS-RWKV and 123.44M for ResViT.Source: C. Comparison with State-of-the-Art… · p. 5
  • Visual comparisons report sharper tissue boundaries and finer anatomical textures, particularly in cortical gray matter and subcortical regions.These regions are identified as places where high-frequency details are critical.Source: C. Comparison with State-of-the-Art… · p. 5

D. Ablation Studies

  • Removing DGCM yields 25.88 dB PSNR and 0.8206 SSIM, while removing SSFA yields 26.02 dB and 0.8270; both modules therefore contribute in the BNU T2w ablation.The backbone alone reaches 25.66 dB and 0.8048; individual PSNR drops total 0.64 dB versus a joint contribution of 0.61 dB, indicating 0.03 dB of functional overlap.Source: D. Ablation Studies · p. 5
  • DGCM removal most affects fine anatomical boundaries, whereas SSFA removal most affects regions with elongated anatomical structures.The ablation describes DGCM as combining edge-aware spatial and frequency-aware channel branches, and SSFA as using strip-pooled global attention and FFT-based spectral modulation.Source: D. Ablation Studies · p. 5

V. DISCUSSION AND CONCLUSION

DGCM-SSFA reports competitive BNU T2w structural fidelity with 11M parameters, while ablations indicate largely orthogonal DGCM and SSFA feature spaces. The discussion identifies slice-wise processing and pathological-case validation as remaining boundaries.

V. DISCUSSION AND CONCLUSION

  • The ablation reports largely orthogonal DGCM and SSFA feature spaces, with only 0.03 dB of functional overlap.The ablation study evaluates the key components on BNU T2w without ensemble.Source: §V · p. 5, Table II · p. 6
  • 0.8431 SSIM on BNU T2w ranks second to FS-RWKV (0.8624), while DGCM-SSFA uses 11M parameters, 24% of FS-RWKV's.The reported comparison describes the model as parameter-efficient while approaching the best benchmark SSIM.Source: §V · p. 5, Table I · p. 6
  • Moderate PSNR reflects a preference for anatomical boundary preservation over pixel-wise intensity accuracy.The discussion characterizes this trade-off as clinically acceptable because diagnostic confidence depends primarily on clear tissue interfaces.Source: §V · pp. 5–6
  • The 2D slice-wise formulation does not exploit inter-slice anatomical consistency; a volumetric 3D extension could improve both fidelity metrics.Source: §V · p. 6
  • The discussion proposes hybrid adversarial, perceptual, and frequency-domain objectives and flags pathological-case validation as a step toward clinical translation.Source: §V · p. 6
Loading 2610.29141v1…