MAGEFormer: geometry-calibrated vision transformers for anisotropic clinical CT segmentation

By embedding physical voxel dimensions and geometric constraints directly into spatial embeddings and attention, MAGEFormer eliminates resampling artifacts and sets new boundary accuracy baselines on abdominal CT.

Primary source: arXiv:2610.04036: https://arxiv.org/abs/2610.04036

What’s new: Vision Transformers (ViTs) deliver impressive volumetric medical image segmentation, but standard formulations rely on an isotropic Euclidean lattice assumption. In clinical practice, computed tomography (CT) scans are inherently anisotropic, often acquired with thick slice intervals relative to fine in-plane pixel spacing. This mismatch creates two distinct problems: spatial misalignment between raw voxel indices and real physical anatomy, and geometric distortion or blur introduced by isotropic resampling preprocessing. In MAGEFormer: Learning Metric-Consistent Representations for Anisotropic CT Segmentation (arXiv:2610.04036, 2 Oct 2026), researchers Jiaying Li and Paolo Remagnino introduce MAGEFormer, a geometry-calibrated transformer architecture that integrates physical metric constraints directly into representation learning.

MAGEFormer replaces naive tokenization and positional embeddings with three targeted mechanisms. First, Metric-Adaptive Spatial Embedding (MASE) dynamically modulates sinusoidal positional frequencies using true physical voxel dimensions, ensuring that spatial distances in feature space reflect physical millimeters. Second, Geometry-Constrained Attention (GCA) enforces spatial distance penalties in the attention matrix, suppressing physically implausible correlations across distant anatomical slices. Third, Geometric View Voting (GVV) aggregates predictions across oriented slice perspectives during inference to minimize discretization bias. When evaluated on the BTCV and FLARE 22 multi-organ abdominal benchmarks against leading CNN and transformer baselines under a unified protocol, MAGEFormer attained superior boundary precision, achieving a 95% Hausdorff Distance (HD95) of 10.58 mm on BTCV and 3.40 mm on FLARE 22 alongside consistent Dice improvements.

Why it matters: Retaining raw anisotropic voxel spacing while maintaining geometric fidelity avoids both the computational overhead and the interpolation artifacts of 3D isotropic resampling. Embedding metric priors directly into attention mechanisms allows clinical vision architectures to process heterogeneous, routine hospital CT scans without specialized re-gridding pipelines.

Caveats: The architecture relies on accurate voxel spacing metadata from DICOM headers, which must be sanitized to avoid corrupt spatial calculations if header tags are missing or improperly scaled. Furthermore, while boundary delineations improve significantly, computational memory scaling for very large volume matrices remains constrained by quadratic attention in high-resolution slices.