Source-linked AI summary
MAGEFormer: Learning Metric-Consistent Representations for Anisotropic CT Segmentation
Jiaying Li, Paolo Remagnino
TL;DR
Anisotropic CT makes voxel indices an imperfect proxy for physical distances, limiting metric consistency in ViT representations. MAGEFormer applies geometry calibration across positional encoding, cross-scale interaction, and inference. It achieves the strongest boundary accuracy among compared baselines on BTCV and FLARE 22, with consistent Dice gains.
Problem
Coarse through-plane spacing makes abdominal CT anisotropic, so voxel-index proximity can differ from physical proximity and hinder boundary-sensitive structures.
Source: §1 · pp. 1–2Method
MAGEFormer enforces metric consistency through spacing-calibrated positional encoding, geometry-constrained cross-scale interaction, and inference-time view aggregation.
Source: §2 · p. 2Results
10.58 mm HD95 on BTCV and 3.40 mm HD95 on FLARE 22 accompany MAGEFormer's strongest boundary accuracy among compared baselines, with consistent Dice gains.
Source: §3.2 · p. 6, §4 · p. 9Takeaways & Limitations
The evaluations support internal geometry calibration as more effective than relying on conventional isotropic preprocessing alone for anisotropic CT segmentation.
Source: §4 · p. 9Most models implicitly assume isotropic geometry after reorientation, while the physical metric remains spacing-dependent.
Source: §2.1 · p. 3
Abstract
from arXiv · showhide
Vision Transformers (ViTs) have shown strong performance in volumetric segmentation, but their effectiveness on clinical CT is limited by an isotropic Euclidean lattice assumption. This conflicts with anisotropic CT acquisition, leading to two key issues: (1) a metric mismatch between voxel indices and physical anatomy, and (2) accuracy degradation from isotropic resampling. To address this, we propose MAGEFormer, a geometry-calibrated framework that embeds physical metric constraints directly into representation learning. Our method introduces Metric-Adaptive Spatial Embedding (MASE) to calibrate positional frequencies using voxel spacing, Geometry-Constrained Attention (GCA) to suppress physically implausible feature correlations, and Geometric View Voting (GVV) to reduce discretization bias during inference. We evaluate MAGEFormer on two multi-organ abdominal CT benchmarks, BTCV and FLARE 22, under a unified protocol against strong CNN and Transformer-based baselines. MAGEFormer achieves the strongest boundary accuracy among the compared methods, with 10.58 mm HD95 on BTCV and 3.40 mm HD95 on FLARE 22, and shows consistent gains in Dice under the same protocol. These results show that geometry-aware internal calibration is more effective than relying on conventional isotropic preprocessing alone for anisotropic CT segmentation.
1 Introduction
Clinical CT is anisotropic, so voxel-index geometry can misrepresent physical distances and hinder boundary-sensitive structures. MAGEFormer calibrates representations across encoding, feature interaction, and inference, achieving the strongest boundary accuracy among the compared baselines.
1 Introduction
- Coarse through-plane spacing makes clinical abdominal CT anisotropic, so equally local voxel neighbors can be physically distant along z and close in the xy-plane.Source: §1 · pp. 1–2
- Existing positional encodings and relative biases operate on voxel indices, while thickness-aware operators do not embed metric constraints in learned representations.Source: §1 · p. 2
- MAGEFormer combines spacing-calibrated positional frequencies, geometry-constrained skip-feature interaction, and sub-voxel view aggregation at inference.MASE reflects physical distance; GCA suppresses physically implausible correlations; GVV reduces discretization bias.Source: §1 · p. 2
- MAGEFormer achieves the strongest boundary accuracy among compared CNN and Transformer-based baselines, with consistent Dice gains on BTCV and FLARE 22.Source: §1 · p. 2
2 Method
MAGEFormer applies metric consistency at positional encoding, cross-scale feature interaction, and inference-time discretization. Its three components operate at complementary stages of a multi-scale U-shaped Transformer pipeline.
2.2 MASE: Metric-Adaptive Spatial Embedding
- MASE calibrates positional encoding using spacing and restricts learnable refinement to the z-axis, the dominant mismatch source in abdominal CT.The refinement interpolates between strict physical calibration and a data-driven optimum.Source: §2 · p. 2, §2.2 · p. 4
2.1 Problem Formulation: Metric Mismatch in Anisotropic Grids
- Equal voxel-index offsets do not imply equal physical displacements, especially along the sparse through-plane axis.At each hierarchy stage, cumulative stride changes the tokens’ effective spacing and metric.Source: §2.1 · p. 3
2.3 GCA: Geometry-Constrained Attention
- GCA constrains encoder–decoder skip interactions with a metric-derived penalty on physically implausible correspondences.Its metric-aware receptive field contracts attention along sparse axes and dilates it along dense ones; large slice spacing suppresses cross-slice attention to the nearest physical neighbors.Source: §2.3 · p. 4
2.4 GVV: Geometric View Voting
- GVV averages predictions from four deterministic sub-voxel views shifted along the anisotropic axis, with shifts and aggregation prescribed by acquisition geometry.Together, MASE, GCA, and GVV address within-scale attention, cross-scale skip interaction, and inference-time discretization, respectively.Source: §2.4 · p. 5
3 Experiments and Results
Under a shared evaluation protocol on BTCV and FLARE 22, MAGEFormer achieves the best overall performance and strongest boundary accuracy among the compared methods. Results also show greater advantages with increasing anisotropy, gains on vulnerable organs, and benefits from its components and inference strategy.
3.1 Experimental Setup
- Evaluation uses 5-fold cross-validation with shared splits on FLARE 22 (50 labeled scans) and BTCV (30 scans), reporting Dice for overlap and HD95 for boundary accuracy.Both datasets annotate 13 organs, with 12 structures in common; inputs are cropped to 96 × 96 × 96.Source: §3.1 · pp. 5–6
3.2 Results
- MAGEFormer achieves the best overall performance on FLARE 22 (93.91% Dice / 3.40 mm HD95) and BTCV (84.11% Dice / 10.58 mm HD95), with the strongest boundary accuracy.Table 2 reports results against nnU-Net, TransUNet, UNETR, Swin-UNETR, and U-Mamba under the unified protocol.Source: Table 2 · p. 6, §3.2 · p. 6
- On BTCV, adding MASE and GCA individually improves the Swin-UNETR baseline, and GVV further enhances both metrics; the full model gains 9.24 Dice points and lowers HD95 to 10.58 mm.Dice rises from 0.7487 to 0.8411, while HD95 falls from 75.21 to 10.58 mm.Source: Table 3 · p. 7, §3.2 · p. 7
- Fig. 2 shows MAGEFormer’s advantage increasing with anisotropy and z-axis downsampling, while it degrades more gracefully than the compared baselines under synthetic degradation.The synthetic downsampling is applied at test time, followed by interpolation back to the original grid.Source: Figure 2 · p. 7, §3.2 · p. 8
- GVV reaches accuracy comparable to standard test-time augmentation while running approximately 4× faster per case in the reported setup.The qualitative comparisons also report smoother, more anatomically consistent boundaries, especially for small organs and thin structures.Source: Figure 2 · p. 7, §3.2 · p. 8
- MAGEFormer has the lowest cross-fold Dice variance on both benchmarks: 0.20 on FLARE 22 and 1.73 on BTCV.Its HD95 standard deviation on FLARE 22 is 0.48.Source: §3.2 · p. 7
- On BTCV, MAGEFormer achieves the best adrenal-gland Dice among compared methods (72.76%) and 82.59% Dice for pancreas segmentation.The paper identifies these organs as especially vulnerable to anisotropic sampling artifacts and partial-volume effects.Source: Table 1 · p. 6, §3.2 · p. 7
- Compared with the listed baselines, MAGEFormer’s qualitative predictions show less multi-label fragmentation and fewer anatomically implausible intra-organ voids.Fig. 3 presents ground truth, five baseline architectures, and MAGEFormer.Source: Figure 3 · p. 8
4 Conclusion
MAGEFormer embeds physical spacing constraints across positional encoding, cross-scale feature interaction, and inference-time view aggregation to address metric mismatch in anisotropic clinical CT. On BTCV and FLARE 22, it achieves the strongest boundary accuracy among compared baselines, with consistent Dice gains.
4 Conclusion
- MAGEFormer addresses the metric mismatch between isotropic Vision Transformers and anisotropic clinical CT through physical spacing constraints in positional encoding, cross-scale interaction, and inference-time view aggregation.The conclusion describes these as embedding constraints directly into the representation-learning and inference pipeline.Source: §4 · pp. 8–9
- MAGEFormer achieves the strongest boundary accuracy among compared baselines on BTCV and FLARE 22, with consistent gains in Dice.The authors conclude that internal geometry calibration is more effective than conventional isotropic preprocessing alone.Source: §4 · p. 9
- Future work will extend geometry-aware calibration to other inherently anisotropic modalities, including multi-parametric MRI.Source: §4 · p. 9