The authors introduce MS-HVED, a sequence-agnostic framework for flexible multi-stack super-resolution of clinical brain MRI that addresses the challenge of anisotropic 2D acquisitions with varying orientations, alignments, and anatomical coverage common in resource-constrained settings. The method independently encodes available axial, coronal, and sagittal stacks into multi-scale Gaussian posteriors using a hetero-orientation variational encoder, then fuses them via a field-of-view-aware Product-of-Gaussians formulation that constrains contributions from unsupported regions. Training employs physics-informed domain randomization to simulate heterogeneous clinical conditions including obliqueness, partial coverage, and imaging artifacts. MS-HVED achieved competitive or superior reconstruction fidelity on synthetic benchmarks and improved downstream tissue segmentation compared to low-resolution inputs. For external clinical validation, the authors conducted a blinded reader study on 44 real cases evaluated by seven expert neuroradiologists, with 30 routine examinations from a Nigerian cohort forming the largest component — representing an under-represented population in medical imaging research. In the Nigerian cohort, MS-HVED was preferred over a supervised baseline for overall image quality in 187 of 210 paired assessments. Across all clinical cohorts, MS-HVED achieved significantly 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).

Read original