GeoSample, a proposed geometry-guided sampling operator, had the best reported mean combination of three segmentation measures among the operators tested in the same three-dimensional U-Net framework: Dice 71.5, HD95 22.8 and ASSD 4.72. The proposed U-Net configuration also had 0.8 million parameters and 108.9G FLOPs, compared with 2.3 million parameters and 194.8G FLOPs for the baseline.
Sampling around local geometry
GeoSample predicts a voxel-wise local three-dimensional rotation frame, uses adaptive step sizes and steers structured symmetric sampling.
The paper presents GeoSample as one formulation spanning feature refinement and resolution-reduction downsampling. Its Consensus Field fuses geometry from an upsampled decoder and an encoder skip connection before skip-connection concatenation.
The comparison covered three kinds of medical images
The experiments used three public datasets: BraTS MRI, MSD HepaticVessel CT and TDSC-ABUS ultrasound. All methods used a fixed-seed split of 75% for training, 10% for validation and 15% for testing.
Under a fixed three-dimensional U-Net macro-architecture, the study compared baseline convolution, Deformable Convolution variants, Dynamic Downsampling and GeoSample, replacing the corresponding operators. Separate plug-in tests used nnU-Net, Swin-UNETR and MedNeXt baseline architectures.
Scores differed across the datasets
On BraTS, the reported baseline-to-proposed pairs were Dice 86.1 to 88.9 and HD95 7.1 to 6.2. On MSD HepaticVessel, the proposed operator was reported as best for Dice and HD95 at 58.2 and 34.3, with near-best ASSD at 7.43.
On TDSC-ABUS, the reported proposed values were Dice 67.4, HD95 27.8 and ASSD 5.78. The paper's account of that comparison emphasized better boundary measures while describing Dice as comparable.
The method was also tested in other networks
Across the tested backbones, the reported plug-in changes ranged from +1.8 to +4.8 for Dice and +2.9 to +6.2 for sensitivity. HD95 changes ranged from -2.5 to -5.2, ASSD changes from -0.09 to -0.49, and specificity was at least 98.2%.
For nnU-Net, the baseline-to-proposed figures were Dice 68.0 to 72.8, parameters 30.8M to 17.4M and FLOPs 1250G to 750G.
Removing components changed the scores
A component-removal test on TDSC-ABUS compared the complete configuration with versions that omitted differentials or the Consensus Field. The complete Dice/HD95/ASSD triplet was 67.4/27.8/5.78; without differentials it was 54.3/51.0/10.71, while without the Consensus Field it was 58.4/48.2/8.52.
The evidence remains computational
The results are computational benchmark, plug-in and ablation evidence from three public datasets and selected model architectures, not direct clinical outcomes or human-intervention evidence. The preprint reports mean segmentation metrics but no inferential tests, confidence intervals, p-values or repeated-run variability estimates.
Case or volume counts and participant-level characteristics are not reported. The evaluation uses fixed-seed splits and a limited set of datasets and backbone configurations, so it does not establish generalization beyond the reported settings.
The work does not show improved patient outcomes, prospective clinical performance or preservation of clinically relevant topology. The paper identifies memory and wall-clock overhead from interpolation-based sampling, together with early-training stability, as issues for future work.
The manuscript is a preprint labeled arXiv:2608.25819v1 and dated 26 Aug 2026. Its abstract states that code is available at the GeoSample repository.
Paper data and sources
Original title: Steer the Sampling, Not the Kernel Grid: Geometry-Guided Sampling Operator for Volumetric Segmentation
Authors: Sizhe Wang, Himashi Peiris, Zhaolin Chen
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text