Brain MRI no longer needs all three stacks: MS-HVED rebuilds from whatever orientations exist, and experts preferred it in 187 of 210 Nigerian pairs
Synopsis
MS-HVED encodes each available axial, coronal or sagittal low-resolution brain MRI stack into a multi-scale Gaussian posterior and fuses them with a field-of-view-aware product, so one trained model accepts any non-empty subset of orientations without retraining. In a blinded reader study on 30 routine brain MRI examinations from a Nigerian cohort, seven expert neuroradiologists preferred MS-HVED over the supervised SegResNet baseline for overall image quality in 187 of 210 paired assessments, and the preference persisted when an axial or coronal input was deliberately omitted.
Interpretation
Routine brain MRI often uses fast 2D multi-slice protocols with thick slices, giving high in-plane but poor through-plane resolution, so clinical protocols commonly acquire axial, coronal and sagittal stacks whose orientation, alignment and coverage differ. MS-HVED lets a single network accept any non-empty subset of those three orientations at inference, without training a separate model per input combination. Earlier learning-based super-resolution methods typically rely on a single under-constrained stack, or on fixed multi-stack configurations that assume idealised acquisition geometry; self-supervised approaches such as TripleSR fuse two orthogonal stacks but need computationally heavy per-subject optimisation and do not directly support a third stack or a missing orientation. On synthetic tests, the same MS-HVED model showed small PSNR differences between the Ax+Sag and Cor+Sag dual-input combinations across the three cohorts, for example 25.67 dB versus 26.42 dB on IXI-T1, indicating similar performance when the available orientation subset changes.
Fusion happens voxel by voxel: an independent encoder turns each orientation into a mean and variance at each scale, forming a local Gaussian posterior, and these are combined into one fused posterior weighted by each stack's field-of-view mask, so regions a stack never covered contribute nothing. This carries the variational fusion idea from hetero-modal segmentation into geometric super-resolution and adds the spatial missingness specific to multi-stack MRI, where not only a whole orientation but also the valid coverage inside each orientation varies with prescription. In the ablation, adding oblique and partial-FOV acquisition simulation raised PSNR over the variant with neither component by 6.06, 5.89 and 4.46 dB on MEN, IXI-T1 and IXI-T2 respectively; adding the FOV masks on top further raised SSIM by 0.014 to 0.049 and lowered perceptual loss by 0.031 to 0.039.
Training volumes come from physics-informed domain randomisation: slice-profile convolution, Fourier-domain cropping for downsampling, random Euler rotations that produce oblique prescriptions, partial-FOV truncation, plus motion ghosting, k-space spikes, bias fields and Rician noise, which allows supervised training without large paired clinical datasets. Earlier domain randomisation focused mainly on contrast, resolution, noise and other appearance or degradation factors, and did not capture the geometry of routine multi-orientation acquisitions, namely oblique prescription and inconsistent coverage across stacks. Models were trained on IXI-T1, IXI-T2 and an internal meningioma cohort and externally validated on 30 routine brain MRI examinations from a Nigerian cohort, which formed the largest component of the 44-case blinded reader study.
On real clinical images, expert readers rated MS-HVED higher than SegResNet for overall quality, and the advantage held when an orientation was withheld. On the paired real cohort the highest PSNR, SSIM and NCC came from the non-learned FOV-weighted stack average; MS-HVED had the lowest perceptual loss there (0.244) and improved over SegResNet on all four reported metrics. MS-HVED was preferred in 187 of 210 Nigerian pairs and 47 of 49 meningioma pairs; by input condition, 152 of 182 full-input pairs, 40 of 42 axial-omitted pairs and 76 of 84 coronal-omitted pairs favoured it.
Perspective
The demonstrated scope is full-reference metrics on synthetic benchmarks, quantitative comparison on a paired real cohort, SynthSeg tissue segmentation on synthetic test sets, and blinded paired ratings by seven neuroradiologists on 44 real cases (30 Nigerian, 7 meningioma, 7 healthy control); the Nigerian and meningioma cohorts have no matched isotropic high-resolution reference, so conclusions there rest on reader ratings rather than full-reference metrics.
The method assumes stacks are rigidly registered and resampled to a common space before fusion, and residual registration error is not explicitly simulated during training, so the fusion mechanism cannot fully correct the inconsistent anatomical evidence it introduces. The clinical meningioma cohort contained only seven cases, a limited range of tumour appearances and protocols. Inter-rater agreement varied by criterion, with median pairwise quadratic-weighted kappa of 0.46 for overall quality and 0.27 for pathology visibility, the latter based only on the seven meningioma cases. The authors state that whether improved reconstructed image quality translates into improved diagnostic interpretation still requires larger multi-institutional studies.
