Flexible Multi-Stack Super-Resolution of Clinical Brain MRI Using Hetero-Orientation Variational Encoding: Expert Reader Evaluation in a Nigerian Cohort
Authors: Kabir Hamzah Muhammad, Lorena Garcia-Foncillas Macias, Soumya Snigdha Kundu, Emily Guilhem, Jon Cleary, Amal Saleh Nour, Yodit Abraham Yaynishet, Chinasa Kalaiwo, Peter Chibuta, Ugumba Kwikima, Udunna Anazodo, Jonathan Shapey, Tom Vercauteren
Organizations: School of Biomedical Engineering & Imaging Sciences, King’s College London, London, United Kingdom · Medical Artificial Intelligence, Crestview Radiology Ltd., Lagos, Nigeria · King’s College Hospital NHS Foundation Trust, London, United Kingdom · Department of Radiology, Guy’s and St. Thomas’ NHS Foundation Trust, London, United Kingdom · School of Medicine, College of Health Sciences, Addis Ababa University, Addis Ababa, Ethiopia · Children’s Hospital of Philadelphia, Philadelphia, United States · African Medical Centre of Excellence, Abuja, Nigeria · University Teaching Hospital, Lusaka, Zambia · Muhimbili Orthopedic Institute & Muhimbili University of Health and Allied Sciences, Dar Es Salaam, Tanzania · Montreal Neurological Institute, McGill University, Montreal, Canada · Department of Neurosurgery, King’s College Hospital NHS Foundation Trust, London, United Kingdom
High-resolution brain MRI is important for neuroradiological assessment, yet routine clinical protocols often rely on fast, anisotropic 2D acquisitions, particularly in resource-constrained settings. These acquisitions may differ in orientation, alignment and anatomical coverage, resulting in partially overlapping image stacks. Existing learning-based super-resolution methods typically rely on single, under-constrained inputs or fixed multi-stack configurations with idealised acquisition geometry. We introduce MS-HVED, a sequence-agnostic framework for flexible multi-stack brain MRI super-resolution. Available axial, coronal and sagittal stacks are independently encoded into multi-scale Gaussian posteriors and fused using a field-of-view-aware Product-of-Gaussians formulation that limits contributions from unsupported regions. Physics-informed domain randomisation simulates heterogeneous clinical acquisition conditions, including obliqueness, partial coverage and imaging artefacts. MS-HVED achieved competitive or superior reconstruction fidelity across synthetic benchmarks and improved downstream tissue segmentation relative to low-resolution inputs. To assess generalisation beyond the training cohorts, we performed external clinical validation on 30 routine brain MRI examinations from a Nigerian cohort, representing an under-represented population in medical imaging research. These examinations formed the largest component of a blinded reader study involving 44 real cases and seven expert neuroradiologists. In the Nigerian cohort, MS-HVED was preferred over the supervised baseline for overall image quality in 187 of 210 paired assessments. Across all clinical cohorts, MS-HVED achieved higher overall-quality scores with all available inputs (5.93 vs. 4.49, p<.001), with the preference persisting when axial or coronal inputs were omitted (p<.001). Code at github.com/cai4cai/ms-hved
Figures & tables
Figure 1: Overview of the proposed MS-HVED framework. Given any non-empty subset of axial, coronal and sagittal anisotropic stacks, each available orientation is processed by an independent variational encoder that predicts latent multi-scale Gaussian posterior parameters. The resulting orientation-specific latent posteriors are fused voxel-wise using a Multi-Scale Product-of-Gaussians, where per-orientation FOV masks modulate each stack’s precision contribution according to its spatial extent. Samples from the fused latent multi-scale posterior are passed through a shared super-resolution decoder to reconstruct the isotropic HR volume. Optional orientation-specific decoders reconstruct the input stacks.
Data
Method
MAE ↓
PSNR ↑
SSIM ↑
Percept. ↓
MEN
ECLARE [ 39 ]
.083 ± .033
16.39 ± 3.8
.57 ± .16
.64 ± .25
TripleSR [ 41 ]
.057 ± .017
19.43 ± 2.4
.68 ± .12
.55 ± .18
SegResNet (Ax)
.035 ± .011
24.02 ± 2.5
.81 ± .08
.17 ± .06
SegResNet (Ax+Cor)
.032 ± .008
24.79 ± 2.3
.81 ± .08
.14 ± .05
SegResNet (Ax+Cor+Sag)
.034 ± .011
24.44 ± 2.5
.82 ± .10
.14 ± .06
MS-HVED (Ax+Cor)
.028 ± .008
25.43 ± 2.3
.88 ± .04
.16 ± .07
Table 1: Synthetic benchmark. All methods are evaluated on the same oblique test inputs, with metrics computed only over the HR target region supported by the stack orientations used by each method. Values report mean ± standard deviation. Best values within each dataset and metric are bolded, and second-best values are underlined.
Data
Orientations
MAE ↓
PSNR ↑
SSIM ↑
Percept. ↓
MEN
Ax+Sag
.027 ± .008
25.69 ± 2.3
.87 ± .04
.16 ± .06
Cor+Sag
.026 ± .007
26.17 ± 2.1
.86 ± .04
.16 ± .06
IXI-T1
Ax+Sag
.027 ± .008
25.67 ± 2.4
.91 ± .04
.15 ± .06
Cor+Sag
.026 ± .008
26.42 ± 2.4
.91 ± .03
.14 ± .06
IXI-T2
Ax+Sag
.024 ± .005
26.50 ± 1.8
.89 ± .05
.12 ± .04
Cor+Sag
.024 ± .005
26.63 ± 1.7
.88 ± .03
.13 ± .04
Table 2: MS-HVED performance for additional dual-orientation input combinations. Both configurations were evaluated using the same trained model for each synthetic benchmark dataset. Values are mean ± standard deviation.
Data
Model
MAE ↓
PSNR ↑
SSIM ↑
Percept. ↓
IXI-T1
DS
.023 ± .007
27.16 ± 2.3
.928 ± .027
.121 ± .049
SA
.024 ± .007
27.36 ± 2.4
.878 ± .075
.110 ± .048
IXI-T2
DS
.021 ± .004
27.71 ± 1.5
.883 ± .057
.102 ± .034
SA
.024 ± .004
27.09 ± 1.3
.825 ± .110
.099 ± .033
MEN
DS
.023 ± .006
27.01 ± 2.1
.883 ± .037
.133 ± .053
SA
.022 ± .006
27.45 ± 2.2
.882 ± .058
.118 ± .046
Table 3: Data-specific (DS) vs sequence-agnostic (SA) MS-HVED Each row reports performance on the held-out test set of the named cohort in the synthetic benchmark. Best values for each cohort and metric are bolded.
Figure 2: Reference HR image and qualitative SR results from the IXI [ 24 ] and MEN synthetic benchmark datasets. The presented results were reconstructed from 5mm anisotropic inputs.
Data
FOV sim.
FOV masks
MAE ↓
PSNR ↑
SSIM ↑
Percept. ↓
MEN
✗
✗
.043 ± .014
20.95 ± 2.76
.821 ± .057
.207 ± .075
✓
✗
.024 ± .007
26.18 ± 2.06
.869 ± .044
.172 ± .071
✓
✓
.023 ± .006
27.01 ± 2.1
.883 ± .037
.133 ± .053
IXI-T1
✗
✗
.045 ± .015
21.27 ± 2.91
.810 ± .087
.195 ± .089
✓
✗
.025 ± .008
26.79 ± 2.57
.879 ± .054
.157 ± .082
✓
✓
.023 ± .007
27.16 ± 2.3
.928 ± .027
.121 ± .049
Table 4: Ablation of oblique partial-FOV simulation and FOV-aware mask conditioning. All variants are evaluated on the same held-out synthetic test sets with oblique partial-FOV geometry. Check marks indicate whether each component is used. Values report mean ± standard deviation.
Method
CSF
GM
WM
Dice ↑
HD 95↓
Dice ↑
HD 95↓
Dice ↑
HD 95↓
MEN
LR Input
0.58±0.20
12.10±28.41
0.69±0.18
9.91±25.56
0.75±0.18
9.67±25.17
SegResNet
0.86±0.05
1.21±0.36
0.90±0.04
1.13±0.35
0.93±0.03
1.18±0.41
MS-HVED
0.87±0.04
1.18±0.26
0.90±0.03
1.06±0.17
0.93±0.02
1.08±0.21
IXI-T1
LR Input
0.60±0.18
7.18±8.67
0.73±0.16
5.47±8.38
0.78±0.16
5.62±8.42
SegResNet
0.87±0.05
1.14±0.20
0.91±0.03
1.05±0.14
0.94±0.02
1.07±0.16
Table 5: Downstream brain tissue segmentation on the synthetic test sets. HR-volume segmentations served as reference. Best values within each cohort and metric are bolded, and second-best values are underlined. Ties at the displayed precision are marked equally.
Method
PSNR ↑
SSIM ↑
NCC ↑
Percept. ↓
NiftyMIC [ 13 ]
15.135 ± 2.90
.374 ± .043
.748 ± .043
–
NiftyMIC(reproduced)
18.82 ± 0.89
.604 ± .115
.702 ± .069
.412 ± .113
ECLARE [ 39 ]
21.88 ± 1.05
.792 ± .014
.622 ± .051
.435 ± .034
TripleSR [ 41 ]
21.56 ± 0.90
.771 ± .014
.584 ± .043
.498 ± .025
Naive mean
22.18 ± 0.89
.797 ± .012
.656 ± .029
.426 ± .016
Naive + FOV
23.35 ± 1.09
.841 ± .009
.750 ± .035
.352 ± .022
Table 6: Full-reference evaluation on the 7 paired real cases. The independently acquired HR images were treated as pseudo-references. Values are mean ± standard deviation. Best and second-best values computed in the present evaluation are bolded and underlined; published NiftyMIC values are included for context and excluded from this ranking.
Figure 3: Qualitative meningioma reconstructions from real low-resolution stacks. Examples show the available anisotropic clinical input stacks and the corresponding Naive+FOV, SegResNet, and MS-HVED reconstructions.
Cohort
Criterion
n
MS-HVED
SegResNet
W/L
p
Nigerian
SNR
210
5.67±2.08
4.08±2.14
184/26
<.001
Sharpness
210
5.32±2.07
4.26±2.14
159/51
<.001
Artefact
210
5.15±1.93
3.30±1.85
199/11
<.001
Consistency
210
5.47±1.92
3.75±1.87
186/24
<.001
Overall quality
210
5.27±1.88
3.58±1.82
187/23
<.001
Meningioma
SNR
49
6.39±1.86
4.37±2.00
47/2
<.001
Table 7: Reader scores stratified by real data cohort. Each reader scored every case and both reconstructions. Scores are mean ± standard deviation. Win ( W ) and Loss ( L ) count case reader pairs in which MS-HVED received a higher or lower score than SegResNet. p -values are from two-sided paired Wilcoxon signed-rank tests.
Input group
Criterion
n
MS-HVED
SegResNet
W/L
p
Full input
SNR
182
6.25±1.97
4.67±2.21
161/21
<.001
Sharpness
182
5.66±1.99
5.09±2.15
100/82
<.001
Artefact
182
5.84±1.88
4.27±2.03
166/16
<.001
Consistency
182
5.97±1.83
4.62±1.93
151/31
<.001
Pathology visibility
49
6.24±1.97
4.76±1.91
44/5
<.001
Overall quality
182
5.93±1.76
4.49±1.94
152/30
<.001
Table 8: Reader scores stratified by reconstruction input condition. Full input denotes use of every real data stack available for a case. In the omitted groups, the named acquired stack was shown to readers but withheld from both methods during inference. Values and tests are as defined as in Table 7 .
Figure 4: Scale-invariant reader preference by real data cohort. For every reader and criterion, net paired preference is 100(W−L)/(W+L) , where W and L are case-level wins and losses for MS-HVED relative to SegResNet. Circles show individual readers, diamonds and horizontal lines show the median and interquartile range. Positive values favour MS-HVED and negative values favour SegResNet. Because the measure uses only the within-pair ordering, it is unchanged by reader-specific shifts or positive rescaling of scores.
Criterion
Weighted κ
Within one point
MAD
SNR
0.34(0.05–0.58)
0.39
2.31
Sharpness
0.30(0.06–0.59)
0.38
2.24
Freedom from artefact
0.39(0.08–0.73)
0.44
2.01
Anatomical consistency
0.37(0.08–0.61)
0.43
1.96
Pathology visibility
0.27(0.06–0.58)
0.38
2.20
Overall quality
0.46(0.08–0.68)
0.47
1.86
Table 9: Inter-rater variability. Values were calculated across shared case–method units. Quadratic-weighted κ is reported as the median and range across reader pairs. Within-one-point agreement and mean absolute difference (MAD) are averaged across reader pairs.
Figure 5: Representative anatomical observations identified during expert review. Examples show local reconstruction errors that were not necessarily reflected by global image-quality metrics. (a) Healthy-control case showing reduced conspicuity of the interpeduncular cistern in both reconstructions. (b) Nigerian case showing reduced distinction between the pituitary gland and clivus in the SegResNet reconstruction. (c) Nigerian case showing increased signal intensity in the genu of the corpus callosum in the SegResNet reconstruction. (d) Nigerian 0.3 T case showing reduced separation of the left ICA flow void from adjacent structures in both reconstructions, more prominently with SegResNet. Arrows indicate the regions discussed by the expert readers.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1: Expert review interface. Screenshot of the blinded multi-planar review interface used for reference-free clinical evaluation. The interface displays registered LR inputs and anonymised reconstructions side by side, allowing expert raters to inspect axial, coronal and sagittal views before assigning image-quality scores.
Figure A.2: Qualitative comparison on the healthy cohort (real data). Reconstructions are shown for a representative case from the paired healthy-control cohort [ 13 ] . Available low resolution (LR) stacks are shown in their inplane and through-plane orientations. Naive+FOV provide non-learned reference baselines obtained by averaging the registered LR stacks with FOV weighting. The separately acquired HR image is shown as a pseudo-reference rather than an absolute ground truth because of contrast differences between the LR and HR acquisitions.
Figure A.3: Qualitative comparison on the Nigerian cohort (real data). Reconstructions are shown for a representative case from the Nigerian clinical cohort [ 51 , 52 ] . Available low resolution (LR) stacks are shown in their inplane and through-plane orientations. Naive+FOV provide non-learned reference baselines obtained by averaging the registered LR stacks with FOV weighting. SegResNet and MS-HVED correspond to the two strongest learned models and are used for the primary expert comparison.
Figure A.4: Qualitative downstream brain-tissue segmentation comparison. A representative SynthSeg [ 6 ] segmentation from an IXI-T1 test case is shown for the LR baseline, SegResNet reconstruction and MS-HVED reconstruction, together with the HR reference segmentation.
Figure A.5: Aggregate paired preference by real data cohort. Bars show the proportions of paired case–reader assessments in which MS-HVED received a higher score or a lower score than SegResNet. Comparisons are made within the same reader and case, so they do not depend on differences in the absolute score ranges used by different readers.
Figure A.6: Mean reader scores for full-input reconstructions from real data. The radar plot is retained as a compact supplementary overview while the box plots in the main text show the underlying distributions. Pathology visibility was scored only for meningioma cases.
Figure A.7: Mean reader scores for deliberately omitted inputs (real data). The axial-omitted group is shown on the left and the coronal-omitted group on the right. These plots summarise the same data whose distributions are presented as box plots in the main text.
Figure A.8: Partial-FOV in real multi-orientation clinical acquisitions. Different anisotropic stacks may cover different anatomical extents.
Figure A.9: MS-HVED reconstruction on different output grids. Representative cases reconstructed at 1mm and 0.5mm isotropic resolution using the same model trained with 1mm targets. The left column shows the original 2D acquisitions at their native in-plane resolution. The 0.5mm reconstructions were generated using sliding-window inference with 128×128×64 patches and 50% overlap. (A) 1.5 T acquisition with 0.5 mm native in-plane resolution and 6.5 mm through-plane spacing, reconstruction grids were 262×290×286 at 1mm and 521×577×569 at 0.5mm . (B) 0.3 T acquisition with 0.5 – 1.0 mm native in-plane resolution and 7.0 – 8.5 mm through-plane spacing, reconstruction grids were 250×249×250 and 497×496×497 , respectively. (C) 1.5 T acquisition with 0.4 mm native in-plane resolution and 6 mm through-plane spacing, reconstruction grids were 292×299×243 and 582×595×483 , respectively.
Medical image super-resolution reconstructs fine spatial detail from low-resolution scans to support clearer anatomy and more useful clinical measurements.
Clinical MRI frequently acquires anisotropic volumes with high in-plane resolution and low through-plane resolution to reduce acquisition time. Multiple orientations are therefore acquired to provide complementary anatomical information. Conventional integration of these views relies on registration followed by interpolation, which can degrade fine structural details. Recent deep learning-based super-resolution (SR) approaches have demonstrated strong performance in enhancing single-view images. However, their clinical reliability is often limited by the need for large-scale training datasets, resulting in increased dependence on cohort-level priors. Self-supervised strategies offer an alternative by learning directly from the target scans. Prior work either neglects the existence of multi-view information or assumes that in-plane information can supervise through-plane reconstruction under the assumption of pre-alignment between images. However, this assumption is rarely satisfied in clinical settings. In this work, we introduce Single-Subject Implicit Multi-View Super-Resolution for MRI (SIMS-MRI), a framework that operates solely on anisotropic multi-view scans from a single patient without requiring pre- or post-processing. Our method combines a multi-resolution hash-encoded implicit representation with learned inter-view alignment to generate a spatially consistent isotropic reconstruction. We validate the SIMS-MRI pipeline on both simulated brain and clinical prostate MRI datasets. Code will be made publicly available for reproducibility: https://github.com/abhshkt/SIMS-MRI
Heejong Kim, Abhishek Thanki, Roel van Herten +2
Department of Radiology, Weill Cornell Medicine, New York, USA · Weill Cornell Graduate School of Medical Sciences, New York, USA · School of Electrical and Computer Engineering, Cornell University and Cornell Tech, New York, USA
Magnetic resonance imaging (MRI) super-resolution is vital for improving diagnostic accessibility, yet most methods treat it as a deterministic mapping from a fixed low-resolution input to a high-resolution target. This overlooks a key property of MRI acquisition physics: spatial resolution and signal-to-noise ratio (SNR) are inherently coupled, making any given low-resolution scan merely one of many possible realizations under varying acquisition trade-offs. We rethink MRI super-resolution as a physics-aware reconstruction problem, in which the goal is to identify the optimal resolution-SNR configuration and then super-resolve it to obtain high-quality MRI results. A key implication of this formulation is that MRI resolution becomes dynamic rather than fixed. To handle such resolution-heterogeneous inputs, we adapt 2D Gaussian Splatting (2D GS) to MRI by formulating reconstruction as a coordinate-based, resolution-agnostic rendering problem. To further enhance fidelity, we introduce three innovations: (1) a prior-aware Gaussian representation that combines an Anatomical Structure Prior for tissue-specific kernel initialization with an Imaging System Prior that captures hardware characteristics via a covariance dictionary; (2) a physics-constrained signal modeling scheme that predicts intrinsic tissue parameters (proton density rho and effective relaxation rate R2) and synthesizes intensities through governing physical equations, ensuring biophysically plausible contrast; and (3) a meta-learning framework that alleviates paired-data scarcity by pretraining on simulated data and adapting to real-world conditions. Extensive experiments on dynamic-resolution datasets and standard benchmarks demonstrate that our method achieves state-of-the-art performance, highlighting its strong potential for clinical deployment.
Lihua Wei, Huatong Gao, Jia Gong +4
School of Biomedical Engineering, ShanghaiTech University, Shanghai 201210, China · School of Biomedical Engineering, ShanghaiTech University, Shanghai 201210, China, and also with the Shanghai Academy of AI for Science, Shanghai, China · Shanghai Academy of AI for Science, Shanghai, China +4
High-resolution Magnetic Resonance Imaging (MRI) is vital for clinical diagnosis but limited by long acquisition times and motion artifacts. Super-resolution (SR) reconstructs low-resolution scans into high-resolution images, yet existing methods are mutually constrained: paired-data methods achieve efficiency only by relying on costly aligned datasets, while implicit neural representation approaches avoid such data needs at the expense of heavy computation. We propose a zero-shot MRI SR framework using explicit Gaussian representation to balance data requirements and efficiency. MRI-tailored Gaussian parameters embed tissue physical properties, reducing learnable parameters while preserving MR signal fidelity. A physics-grounded volume rendering strategy models MRI signal formation via normalized Gaussian aggregation. Additionally, a brick-based order-independent rasterization scheme enables highly parallel 3D computation, lowering training and inference costs. Experiments on two public MRI datasets show superior reconstruction quality and efficiency, demonstrating the method's potential for clinical MRI SR.
Shuting Liu, Lei Zhang, Wei Huang +2
School of Artificial Intelligence, Sichuan University, Chengdu 610065, China · Institute of High Performance Computing, A*STAR, Singapore 138632