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

1. Introduction

• Full-image losses underweight sparse interfaces; boundary losses cannot encode continuous tissue mixtures, and multi-contrast guidance requires extra acquisition, leaving fixed mixture-weight training underexplored (Lyu et al., 2020).PVE modeling has primarily focused on segmentation and tissue-fraction estimation (Zhang et al., 2001; Van Leemput et al., 2003; Tohka et al., 2004).Source: §1 · pp. 1–2
• AGW-PBR converts tissue-mixture entropy from fixed, quality-controlled PVE sidecars into mean-normalized spatial weights that emphasize transitions while retaining full-image supervision.The sidecars are training resources, not prediction targets; inference requires only the LR image.Source: §1 · p. 2
• The study establishes quality control for PVE sidecars and uses independent SynthSeg segmentations to assess tissue-interface and non-interface regions on T2-weighted IXI images.Source: §1 · p. 2
• The LR-only AGW backbone combines LR-derived structural guidance, compact soft latent assignment, and bounded grid-anchored residual warping.“Gaussian-parameter” refers to the warping module’s displacement and gating parameterization.Source: §1 · p. 2
• AGW-PBR is evaluated on IXI at 2×, 4×, and 6× with common splits, degradation settings, and subject-level paired analysis; AGW is separately evaluated on fastMRI without PVE sidecars.Source: §1 · p. 2

2.1. MRI Super-Resolution Architectures

• MRI SR progressed from interpolation and handcrafted priors to networks modeling local and global context, queryable continuous representations for arbitrary-scale 3-D reconstruction, and diffusion refinement for arbitrary-scale in-plane reconstruction (Keys, 1981; Stark and Oskoui, 1989; Zhang et al., 2018a,b; Liang et al., 2021; Wu et al., 2023; Wang et al., 2026).ArSSR learns an implicit representation for arbitrary-scale 3-D reconstruction; PRDDiff progressively recovers high-frequency k-space components for arbitrary-scale in-plane reconstruction.Source: §2.1 · p. 2

2.2. Anatomical, Segmentation, and Tissue-Aware Guidance

• Anatomical guidance ranges from additional MR contrasts to learned structure maps, segmentation constraints, and internally predicted categorical tissue cues (Lyu et al., 2020).Multi-contrast methods require acquisition and registration; categorical per-voxel probabilities differ from continuous tissue fractions that express multiple tissues’ relative contributions within a voxel.Source: §2.2 · pp. 2–3

2.3. Partial-Volume Effects and Tissue-Mixture Modeling

• At CSF, GM, and WM interfaces, partial-volume intensities reflect tissue composition and imaging conditions, while continuous tissue fractions characterize voxel mixtures beyond binary boundaries.This study derives training-time spatial reconstruction weights from fixed, quality-controlled CSF/GM/WM PVE pseudo-label sidecars and uses independent SynthSeg segmentations for primary test-time regional masks; estimated fractions depend on image quality and modeling assumptions (Zhang et al., 2001; Van Leemput et al., 2003; Tohka et al., 2004).Source: §2.3 · p. 3

3.1. Problem Formulation and Method Overview

• AGW-PBR separates LR-only reconstruction from training-only PVE-balanced supervision, using Sobel guidance, soft latent assignment, and bounded grid-anchored residual warping.Fixed offline PVE sidecars weight reconstruction errors within quality-controlled support but never enter the network.Source: §3.1 · p. 3

3.2. Anatomy-Guided Feature Extraction

• The feature extractor derives Sobel-gradient magnitude from the LR image to modulate learned features, and this image-derived guidance is active during training and inference.Because the gradient is computed from the LR image, it is neither a tissue segmentation nor an external anatomical label.Source: §3.2 · p. 3

3.3. Compact Latent Gaussian-Parameter Dictionary

• A learnable temperature-scaled soft assignment mixes four learned bases into a five-dimensional local code, which a parameter-residual branch refines.The basis bank is jointly optimized with the network.Source: §3.3 · p. 4

3.4. Grid-Anchored Residual Warping and Reconstruction

• The five local-code components are converted into bounded displacements and a residual-intensity gate that define shifted sampling coordinates for residual reconstruction.The displacement-amplitude gates modulate the shifts, while the residual gate controls the predicted residual.Source: §3.4 · p. 4

3.5. PVE-Balanced Reconstruction Objective

• Fixed, quality-controlled CSF, GM, and WM fraction maps yield tissue-mixture entropy and a valid-PVE mask for unit-mean weighted L1 reconstruction during training.The PBR term is computed over valid support; the sidecars are offline resources rather than trainable predictions.Source: §3.5 · p. 5

3.6. Inference Boundary

• At inference, the model uses only the LR image, with Sobel guidance and target-grid coordinates generated internally; PVE sidecars are not required.SynthSeg-derived interfaces are used solely to construct offline test-set evaluation masks.Source: §3.6 · p. 5

4.1. Experimental Settings

• IXI was the primary benchmark, with subject-level splits of 329 training, 47 validation, and 99 test subjects; fastMRI provided external evaluation on 200, 30, and 43 subjects, respectively.Full-image performance was evaluated at all three IXI scales, while independent SynthSeg regional analysis was conducted at 4×.Source: §4 · p. 5, §4.1 · p. 6

4.2. Degradation Protocol and Implementation Details

• LR images were synthesized by retaining the central 1/s × 1/s k-space region; IXI training used Charbonnier plus PBR loss, whereas fastMRI used Charbonnier loss alone.The retained k-space fractions were 25%, 6.25%, and 2.78% at 2×, 4×, and 6×, respectively; training ran for 80 epochs.Source: §4.2 · p. 6

4.3. Quality Control of the PVE Pseudo-label Sidecars

• Quality-controlled FAST-derived CSF, GM, and WM fractions from 491 IXI subjects yielded valid PVE sidecars for 475 subjects; independent test-only SynthSeg interfaces supported 4× regional evaluation.Regional metrics used interface and non-interface intracranial masks; paired comparisons were subject-level, with bootstrap confidence intervals and Holm-corrected Wilcoxon tests.Source: §4.3 · pp. 6–7

4.4. Results on the IXI Dataset

• On IXI, AGW-PBR achieved the highest PSNR and SSIM at 2×, 4×, and 6×, with positive paired differences versus all compared methods and strongest practical gains at higher factors.At 4×, it also achieved the highest PSNR and SSIM in independent SynthSeg-defined interface and non-interface intracranial regions.Source: §4.4 · p. 8, Table 4 · p. 10

4.5. Results on the fastMRI Dataset

• On fastMRI at 4×, the PVE-free AGW-backbone achieved the highest PSNR and SSIM among compared methods.The backbone was trained from scratch using conventional full-image supervision, without PVE weighting or tissue-interface supervision.Source: §4.5 · p. 8, Table 6 · p. 10

4.6. Ablation Study

• Ablations showed that valid-support supervision improved on the Charbonnier-only backbone, entropy modulation added gains, spatially shuffled or random weights underperformed, and soft assignment beat Hard-ST routing.A1 outperformed A0, while A2 further improved metrics—especially interface fidelity—over A1; A3 and A4 remained below A2, and A5 returned to the A0 range. These comparisons support the contribution of valid support, anatomically aligned entropy weighting, and soft latent assignment.Source: §4.6 · pp. 10–11, Table 7 · p. 11

4.7. Application on Brain Tumor Segmentation Task

• On BraTS tumor segmentation after direct transfer from IXI, AGW-PBR achieved competitive WT Dice, the lowest HD95, and improved both metrics over AGW-backbone.A frozen probe evaluated whole-tumor masks on the 240-subject test cohort; AGW-PBR also produced more spatially coherent qualitative predictions.Source: §4.7 · pp. 11–12, Table 8 · p. 11

5. Discussion

• AGW-PBR achieves the strongest full-image PSNR and SSIM across IXI scales and highest fidelity in SynthSeg interface and non-interface regions.Subject-level paired differences are positive against all evaluated comparators.Source: §5 · p. 12
• A1 improves over the Charbonnier-only A0, and A2 further improves under the same valid-support protocol, with stronger gains in SynthSeg interface fidelity.The comparison indicates that validated PVE-support supervision helps reconstruction and entropy weighting directs optimization toward anatomical transitions.Source: §5 · p. 12
• Shuffled-entropy A3 and random-field A4 both decline from A2, showing that nonuniform weighting without anatomical spatial correspondence cannot reproduce its gains.Both controls preserve the architecture, valid support, normalization, and PBR reduction used by A2.Source: §5 · p. 12
• Hard-ST A5 remains close to A0 and below A2, favoring soft latent-basis assignment that combines multiple basis vectors at each spatial location.The basis vectors are latent components, not tissue-specific representations.Source: §5 · p. 12
• Without PVE sidecars, PBR supervision, or IXI initialization, the AGW-backbone trained from scratch achieves the strongest full-image performance among evaluated fastMRI methods.Its margin over the strongest competitor is modest; this experiment evaluates the LR-only backbone under an external SR-only protocol.Source: §5 · p. 12
• Training entropy comes from quality-controlled FAST tissue fractions, while test-only SynthSeg masks independently define IXI regions; subject-level metrics are aggregated from slice averages.All methods share a subject–slice manifest, and the entropy–interface analysis assesses spatial agreement between FAST-derived entropy and SynthSeg interfaces.Source: §5 · p. 12
• Evaluation is limited by synthetic central k-space truncation, 2D processing, absent paired clinical LR/HR acquisitions and reader assessments, and regional SynthSeg analysis limited to IXI.Future work should test independent cohorts, scanners, protocols, realistic LR/HR pairs, 3D consistency, reader studies, downstream tasks, and anatomy-specific resources for nonbrain MRI; fastMRI tested only the PVE-free backbone.Source: §5 · p. 13

6. Conclusion

• AGW-PBR achieves the strongest full-image performance across evaluated IXI settings and consistent performance in independently defined SynthSeg interface and non-interface regions.It combines an LR-only anatomy-guided warping pathway with a training-time PVE-balanced objective derived from quality-controlled tissue-mixture entropy within validated PVE support.Source: §6 · p. 13
• Ablations support valid-PVE-support supervision, spatially aligned entropy weighting, and soft latent-basis assignment.Source: §6 · p. 13
• The independently trained AGW-backbone performs strongly on fastMRI without PVE sidecars or PBR supervision, while retaining LR-only inference.Source: §6 · p. 13

• Xiao Tong led conceptualization, methodology, software, validation, formal analysis, and the original draft; Wenyun Yang contributed conceptualization, supervision, and review and editing.Source: CRediT authorship contribution statement · p. 13
• Ziheng Zhang handled resources and data curation, while Jingzhi Han contributed validation and investigation.Source: CRediT authorship contribution statement · p. 13
• Zhaochu Luo contributed supervision, funding acquisition, and review and editing; Jinbo Yang contributed supervision, funding acquisition, and project administration.Source: CRediT authorship contribution statement · p. 13

Declaration of competing interest

• Ziheng Zhang is employed by Beijing MagnVue Medix Co., Ltd., while the remaining authors declare no known competing financial interests or personal relationships that could have influenced the work.Source: Declaration of competing interest · p. 13

• IXI and fastMRI are publicly available from their repositories under applicable data-use conditions, but the authors are not authorized to redistribute the original data.Source: Data availability · p. 13
• 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.Source: Data availability · p. 13