Tissue-Mixture Entropy-Weighted Reconstruction for Partial-Volume-Aware Brain MRI Super-Resolution

Xiao Tong; Wenyun Yang; Ziheng Zhang; Jingzhi Han; Zhaochu Luo; Jinbo Yang

Source-linked AI summary

Tissue-Mixture Entropy-Weighted Reconstruction for Partial-Volume-Aware Brain MRI Super-Resolution

Xiao Tong, Wenyun Yang, Ziheng Zhang, Jingzhi Han, Zhaochu Luo, Jinbo Yang

arXiv:2608.26647v1Aug 27, 2026cs.CV

TL;DR

Brain MRI super-resolution must reconstruct fine tissue transitions from incomplete low-resolution observations, while full-image losses can underemphasize these regions. AGW-PBR combines an LR-only anatomy-guided backbone with training-time PVE entropy weighting, and it improves full-image IXI performance and regional fidelity while retaining strong PVE-free fastMRI performance.

Problem
MRI super-resolution is ill-posed, and fine structures and tissue transitions are difficult to reconstruct because high-frequency information is removed and PVE regions occupy limited image area.
Method
AGW-PBR combines an LR-only backbone using structural guidance, soft latent assignment, and bounded residual warping with entropy-based spatial reconstruction weights from fixed PVE sidecars.
Results
AGW-PBR achieves the strongest full-image IXI performance across evaluated settings and consistent regional performance, while the AGW backbone performs strongly on fastMRI without PVE supervision.
Takeaways & Limitations
PVE-balanced training improves brain T2-weighted MRI super-resolution while retaining LR-only inference.
Takeaways & Limitations
The study uses synthetic central k-space truncation, 2D slices, no paired clinical LR/HR acquisitions or reader assessments, and limited regional and fastMRI evaluations.

Abstract

from arXiv · show

Full-image objectives in brain magnetic resonance imaging (MRI) super-resolution (SR) can underweight tissue-transition regions affected by the partial-volume effect (PVE), as these regions occupy only a small fraction of the image. Binary boundaries also do not capture the continuous mixture of cerebrospinal fluid, gray matter, and white matter within a voxel. We propose Anatomy-Guided Gaussian-Parameter Warping with PVE-Balanced Reconstruction (AGW-PBR), which combines a low-resolution (LR)-only reconstruction backbone with a training-time objective that emphasizes tissue transitions. The backbone integrates LR-derived Sobel guidance, soft latent-basis assignment, and bounded grid-anchored residual warping. Fixed, quality-controlled tissue fractions derived from registered T1/T2/PD IXI images are converted into tissue-mixture entropy, which defines mean-normalized reconstruction weights within validated PVE support. These sidecars are used only during training, and inference requires only the LR image. AGW-PBR is evaluated on T2-weighted IXI images at 2x, 4x, and 6x using three seeds and subject-level paired analyses. At 4x, test-only SynthSeg masks independently assess reconstruction in tissue-interface and non-interface regions. Targeted ablations examine valid-support supervision, spatially aligned entropy weighting, and soft latent assignment. The AGW-backbone is also trained from scratch on fastMRI at 4x without PVE supervision. AGW-PBR improves full-image reconstruction across the tested IXI scales and regional fidelity at 4x, while the PVE-free backbone retains strong performance on fastMRI. These findings support tissue-mixture entropy weighting for partial-volume-aware brain MRI SR.

1. Introduction

Brain MRI super-resolution can recover high-resolution images from low-resolution observations, but tissue transitions are difficult because partial-volume errors occupy little of the image. AGW-PBR addresses this gap with an LR-only anatomy-guided backbone and entropy-weighted training objective, evaluated across IXI scales and regional settings.

Motivation: Higher-resolution brain MRI improves delineation of cortical folds, tissue interfaces, and small structures but increases scan time and acquisition trade-offs.Super-resolution offers a computational alternative that does not require scanner hardware changes.
Motivation: MRI super-resolution is ill-posed because high-frequency information removed during acquisition cannot be uniquely recovered from the LR observation.The reconstruction challenge includes preserving fine structures and tissue transitions supported by the observation.
Motivation: PVE-affected CSF, GM, and WM interfaces occupy a small image fraction, so their errors can have limited influence under full-image objectives dominated by homogeneous regions.Binary boundaries also provide less information than continuous tissue mixtures, motivating tissue-mixture weighting.
Proposed approach: AGW-PBR combines LR-derived structural guidance, soft latent assignment, and bounded grid-anchored residual warping with training-time entropy-based spatial reconstruction weights.Fixed, quality-controlled PVE sidecars define the weights, while inference uses only the LR image.
Evaluation: The study evaluates IXI T2-weighted SR at 2×, 4×, and 6× with subject-level paired analyses, independent SynthSeg regional masks, and a separate fastMRI backbone experiment.The fastMRI protocol excludes PVE sidecars and evaluates the reconstruction backbone under an independent SR-only setting.

2. Related Work

MRI super-resolution research spans convolutional, Transformer, continuous, diffusion, and anatomy-guided approaches. This work distinguishes continuous tissue-fraction guidance from binary or categorical cues by using fixed PVE sidecars to derive training-time entropy weights and independent regional evaluation masks.

Representation-based methods: Convolutional and Transformer architectures provide local, multiscale, and longer-range feature representations for image-domain reconstruction.Transformers use attention to capture longer-range dependencies.
Continuous representations: Continuous MRI SR representations predict intensities at queried coordinates, supporting flexible output grids and arbitrary-scale or inter-slice reconstruction.Examples include implicit representations such as ArSSR and SA-INR.
Diffusion methods: Diffusion-based MRI SR reconstructs high-resolution images through conditional generative refinement and has been adapted to reduce sampling costs or recover high-frequency k-space components.The cited MRI approaches include residual-shifting, partial-diffusion, and progressive k-space recovery strategies.
Anatomical guidance: Anatomical guidance has been incorporated through additional MR contrasts, cross-modality feature alignment, deformable attention, learned anatomical representations, and segmentation constraints.These approaches use complementary anatomical information or task-specific structure constraints during reconstruction.
Tissue-aware guidance: Internal tissue guidance may use categorical probabilities, whereas continuous CSF/GM/WM fractions describe relative multi-tissue composition within a voxel.A boundary marks where a transition occurs, while tissue fractions characterize the mixture at that location.
Position of this work: This study uses fixed, quality-controlled PVE pseudo-label sidecars for entropy-based training weights and independent SynthSeg segmentations for primary test-time regional masks.The design targets tissue-mixture weighting rather than internally predicted categorical tissue cues.

3. Methods

AGW-PBR uses an LR-only reconstruction pathway combining Sobel-guided features, soft latent-basis assignment, and bounded grid-anchored residual warping. During training, fixed quality-controlled PVE sidecars convert tissue-mixture entropy into normalized reconstruction weights over valid support, while inference uses only the LR image.

3.1. Problem Formulation and Method Overview: The problem formulation reconstructs ISR from ILR and a fixed degradation-and-resampling operator D using network Fϕ.ILR, IHR, and ISR denote the LR-derived observation, HR target, and reconstructed output on the target grid.
3.2. Anatomy-Guided Feature Extraction: The LR-only pathway derives Sobel-gradient guidance, modulates base features, predicts a five-dimensional local code, and uses it for bounded residual warping.The image-derived guidance remains active during both training and inference.
3.2. Anatomy-Guided Feature Extraction: Sobel kernels produce gradient response fields from ILR, and the resulting magnitude is used as image-derived guidance rather than tissue segmentation.The numerical stabilizer is ϵg = 10^-6.
3.3. Compact Latent Gaussian-Parameter Dictionary: Soft assignment normalizes K basis logits along the basis axis, while a learned basis bank and parameter-residual operator form the local code.AGW-PBR uses K = 4 and Dη = 5; the temperature τ is learnable and initialized to 1.
3.4. Grid-Anchored Residual Warping and Reconstruction: The local code is decoded into bounded displacements and a residual-intensity gate that define the shifted sampling grid eu, whose sampled fields are combined with ILR to produce ISR.The warp coefficient is λwarp = 0.05, with bicubic residual sampling, bilinear gate sampling, reflection padding, and aligned corner pixels.
3.5. PVE-Balanced Reconstruction Objective: Fixed CSF, GM, and WM PVE fractions are constructed offline, converted to clipped entropy, and used within validated support to create unit-mean weighted PBR supervision.An empty valid support produces a safe zero PBR contribution, and sidecars do not receive gradients or enter inference.

4. Experiments

Experiments evaluate AGW-PBR on IXI across three upsampling scales, with independent regional analysis at 4×, and assess the PVE-free backbone on fastMRI. AGW-PBR improves global and interface reconstruction, while ablations support valid-support supervision, aligned entropy weighting, and soft assignment.

IXI results: AGW-PBR achieves the highest PSNR and SSIM across IXI upsampling factors, with clearer advantages at 4× and 6× than at 2×.The paired analysis reports positive differences against all compared methods, with stronger practical gains under higher upsampling.
Regional fidelity: AGW-PBR achieves the highest PSNR and SSIM in both independent SynthSeg-defined interface and non-interface regions at 4×.The regional evaluation assesses whether interface gains occur without reduced fidelity elsewhere in the intracranial volume.
Overall findings: The IXI evidence is consistent across qualitative inspection, global metrics, paired analysis, and PVE-aware regional evaluation, with stronger benefits under more challenging degradation.The study reports modest but stable improvements under mild degradation and more pronounced advantages as downsampling increases.
fastMRI results: AGW-backbone achieves the highest PSNR and SSIM among compared fastMRI methods under conventional full-image supervision at 4×.Qualitative fastMRI results also report the lowest error-map intensity in highlighted regions.
Ablation study: Valid-support supervision improves the backbone, entropy modulation adds interface gains, anatomical alignment preserves those gains, and soft assignment outperforms Hard-ST routing.A1 exceeds A0, A2 improves over A1, shuffled or random entropy guides remain below A2, and A5 remains below A2.

5. Discussion

AGW-PBR improves both overall and anatomically localized IXI reconstruction, with ablations supporting valid-support supervision, aligned entropy weighting, and soft latent assignment. The study also identifies boundaries in data realism, dimensionality, and regional validation.

AGW-PBR achieves the strongest full-image PSNR and SSIM across evaluated IXI scales and positive paired differences against all comparators.
AGW-PBR also achieves the highest fidelity in independently defined SynthSeg interface and non-interface regions.These regional evaluations extend the comparison beyond full-image agreement.
Valid-support supervision improves both full-image and tissue-interface fidelity, while entropy weighting further improves performance under the same mask and protocol.The stronger gain in the independent SynthSeg interface region indicates that entropy weighting directs optimization toward anatomical transitions.
Shuffled-entropy and random-field controls underperform A2, showing that nonuniform weighting requires anatomical spatial correspondence.Both controls retain the architecture, valid support, normalization, and PBR reduction used by A2.
Soft latent assignment outperforms Hard-ST routing under the same validated entropy weighting and PBR objective.Soft assignment preserves mixed latent representations at each spatial location, whereas Hard-ST selects one basis vector in the forward representation.
The independently trained AGW-backbone achieves the strongest full-image fastMRI performance without PVE sidecars or PBR supervision, but its margin over the strongest competitor is modest.The study uses synthetic central k-space truncation, two-dimensional slices, no paired clinical LR/HR acquisitions or reader assessments, and IXI-limited regional evaluation.
Future validation should include independent cohorts, scanners, acquisition protocols, realistic LR/HR pairs, three-dimensional reconstruction, reader studies, and downstream tasks.Extending beyond brain MRI would require anatomy-specific tissue-mixture definitions and separately validated auxiliary resources.

6. Conclusion

AGW-PBR combines an LR-only anatomy-guided backbone with training-time PVE-balanced weighting and performs strongly across IXI reconstruction and regional evaluations. Ablations support its principal design choices, while the backbone also performs strongly on fastMRI without PVE supervision.

AGW-PBR combines LR-only anatomy-guided warping with a training-time objective based on quality-controlled tissue-mixture entropy within validated PVE support.
AGW-PBR achieves the strongest full-image performance across evaluated IXI settings and consistent performance in SynthSeg interface and non-interface regions.
Ablations support valid PVE-support supervision, spatially aligned entropy weighting, and soft latent-basis assignment.
The independently trained AGW-backbone performs strongly on fastMRI without PVE sidecars or PBR supervision, while inference remains LR-only.

CRediT authorship contribution statement

The CRediT statement assigns contributions across conceptualization, methodology, software, validation, analysis, writing, supervision, resources, funding, and project administration.

Xiao Tong contributed conceptualization, methodology, software, validation, formal analysis, and original-draft writing.
Wenyun Yang contributed conceptualization, supervision, and writing-review and editing.
Ziheng Zhang contributed resources and data curation, while Jingzhi Han contributed validation and investigation.
Zhaochu Luo contributed supervision, funding acquisition, and writing-review and editing; Jinbo Yang contributed supervision, funding acquisition, and project administration.

Declaration of generative AI use

The authors used Chat-GPT for language polishing and grammatical corrections, then reviewed and edited the manuscript and accepted responsibility for its content.

Chat-GPT assisted with language polishing and grammatical corrections during article preparation.
The authors reviewed and edited the manuscript after using the tool and took full responsibility for the published content.

Declaration of competing interest

The paper reports one author’s employment at Beijing MagnVue Medix Co., Ltd. and states that the remaining authors declared no known competing interests or relevant personal relationships.

Ziheng Zhang is employed by Beijing MagnVue Medix Co., Ltd.
The remaining authors declare no known competing financial interests.
The remaining authors also declare no personal relationships that could have influenced the reported work.

Data availability

The IXI and fastMRI datasets are publicly available from their respective repositories under applicable data-use conditions, while the authors cannot redistribute the original data. Study materials and derived results are available from the corresponding author upon reasonable request.

The IXI and fastMRI datasets are publicly available from their respective repositories, subject to applicable data-use conditions.
The authors are not authorized to redistribute the original IXI and fastMRI data.
Subject-level split manifests, preprocessing and quality-control scripts, model configurations, evaluation code, and derived subject-level results are available from the corresponding author upon reasonable request.