Flexible Multi-Stack Super-Resolution of Clinical Brain MRI Using Hetero-Orientation Variational Encoding: Expert Reader Evaluation in a Nigerian Cohort

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 VercauterenarXiv 2026 · arXiv

TL;DR

MS-HVED is a new deep learning framework that reconstructs high-resolution, isotropic 3D brain MRI volumes from any combination of standard 2D clinical scans (axial, coronal, or sagittal). By using a "Product-of-Gaussians" fusion strategy and training with realistic clinical simulations, it handles missing orientations and partial anatomical coverage far better than existing methods, earning strong preference from expert neuroradiologists in a first-of-its-kind evaluation on a Nigerian clinical cohort.

What Problem Does This Paper Address?

In an ideal world, every brain MRI would be a high-resolution, isotropic 3D volume. In reality, especially in resource-constrained settings, clinical protocols rely on fast 2D acquisitions. These scans have high resolution within each slice but are very "thick" (anisotropic), leading to blurry images when viewed from other angles.

To compensate, radiologists often acquire multiple "stacks" in different orientations (axial, coronal, sagittal). However, existing super-resolution (SR) algorithms struggle with the "messiness" of real-world clinical data:

  1. Rigidity: Many models require a fixed set of inputs (e.g., always needing all three orientations).
  2. Partial Coverage: Different stacks might cover different parts of the brain or be prescribed at tilted (oblique) angles.
  3. Domain Gap: Models trained on "perfect" synthetic data often fail when faced with real-world clinical artifacts like motion, noise, and varying contrast.

Methods

The authors introduce MS-HVED (Multi-Stack Hetero-Orientation Variational Encoding), a framework designed for flexibility and robustness.

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.

Key technical innovations include:

  • Variational Hetero-Orientation Fusion: Each available orientation is processed by its own encoder, which predicts a latent Gaussian posterior (mean and variance). This allows the model to learn its own "confidence" for different regions of the image.
  • FOV-Aware Product-of-Gaussians (PoG): The latent representations are fused using a PoG formulation. Crucially, this is "Field-of-View (FOV) aware"—it uses masks to ensure that regions not covered by a specific scan don't contribute noise to the final fusion.
  • Physics-Informed Domain Randomization: Instead of relying on scarce paired clinical data, the authors trained the model on high-resolution data degraded by a sophisticated forward model. This model simulates clinical "messiness," including oblique slice angles, partial coverage, motion ghosting, and Rician noise.
  • Spectrally Normalised Architecture: The network avoids standard activation normalization (like Batch Norm) to preserve absolute intensity relationships, using spectral normalization instead to maintain training stability.

Datasets

The study utilized a mix of public and internal datasets for a comprehensive evaluation:

  • Training/Synthetic Benchmarks: The public IXI dataset (T1 and T2) and an internal meningioma (MEN) cohort.
  • Paired Real Data: A small cohort of 7 healthy controls with matched low-res and high-res scans.
  • Clinical Generalization: 7 clinical meningioma cases from the authors' institution, often featuring missing orientations.
  • Nigerian Cohort: 30 routine clinical examinations from a public Nigerian dataset. This is a critical addition, as African populations are significantly under-represented in medical imaging research.

Key Insights and Ideas

  • Spatial "Missingness" Matters: The authors realized that "missing data" isn't just about a missing orientation; it's spatial. A sagittal scan might cover the whole brain, while an axial scan only covers the mid-brain. The FOV-aware fusion treats this spatial variation as a first-class citizen.
  • Variational Fusion as a Weighted Average: By predicting a variance (uncertainty) for every voxel, the model naturally performs a precision-weighted fusion. It "trusts" the high-fidelity in-plane information from one stack more than the blurry through-plane information from another.
  • Sequence Agnosticism: A single model trained on both T1 and T2 data performed competitively with models trained on specific sequences, suggesting the framework learns geometric features that transcend specific MRI contrasts.

Key Results

  • Expert Preference: In a blinded study with seven neuroradiologists, MS-HVED was overwhelmingly preferred. In the Nigerian cohort, it was chosen over the strongest baseline (SegResNet) in 187 out of 210 paired assessments.
  • Clinical Quality: Across all clinical cohorts, MS-HVED achieved significantly higher overall quality scores (5.93 vs. 4.49, $p < .001$).
  • Robustness to Missing Data: The preference for MS-HVED remained strong even when the researchers deliberately withheld axial or coronal stacks, proving the model's flexibility.
  • Downstream Utility: Reconstructions improved the performance of automated tissue segmentation tools (like SynthSeg) compared to using raw low-resolution inputs.

Limitations and Future Work

  • Registration Dependency: The model assumes that different stacks have been pre-registered. While it handles partial overlap, it does not yet explicitly model or correct for residual registration errors (e.g., patient motion between scans).
  • Low-Field MRI: Evaluation on 0.3T scans (very low field) showed some specific artifacts, such as reduced separation in flow voids, indicating a need for more dedicated training data for ultra-low-field scanners.
  • Pathology Diversity: While the model handled meningiomas well, further validation is needed across a wider variety of brain pathologies and lesions.

Summary updated

Abstract

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
BibTeX
@misc{muhammad2026flexible,
  author = {Kabir Hamzah Muhammad and Lorena Garcia-Foncillas Macias and Soumya Snigdha Kundu and Emily Guilhem and Jon Cleary and Amal Saleh Nour and Yodit Abraham Yaynishet and Chinasa Kalaiwo and Peter Chibuta and Ugumba Kwikima and Udunna Anazodo and Jonathan Shapey and Tom Vercauteren},
  title = {{Flexible Multi-Stack Super-Resolution of Clinical Brain MRI Using Hetero-Orientation Variational Encoding: Expert Reader Evaluation in a Nigerian Cohort}},
  year = {2026},
  eprint = {2611.21412},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21412},
}

Loading related content...

Figure