When voxels are not equal: spacing-driven metric bias and metric-uncertainty decoupling in segmentation

Conference item

CT slice thickness varies substantially within and across clinical datasets, yet segmentation evaluation conventions treat every voxel as equally informative regardless of its physical size. We examine three consequences of this for KiTS23 kidney/tumour/cyst segmentation (283 cases across four cross-validation folds, z-spacing 0.5–5.0 mm) with a trained nnU-Net model [8]. First, spacing-naive boundary metrics (Symmetric Boundary Dice, cubic neighbourhood) artificially inflate scores for thick-slice cases by +0.031 Dice points, while a physically-defined neighbourhood variant reverses this to −0.012, confirmed across 283 properlysplit cases. Second, we show that three internally consistent but conventionally conflated ways of aggregating Dice disagree by up to 0.25 Dice points for some structures, and reverse direction depending on structure. Third, model z-boundary entropy increases significantly with z-spacing (Spearman ρ = 0.233, p < 0.001) in the opposite direction to what simple information loss would predict, consistent with the hypothesis that the model has learned to reflect annotation imprecision beyond image ambiguity. Notably, metric inflation and this uncertainty signal show only a weak association at the case level (ρ = 0.083, p = 0.166), consistent with the two effects operating through largely independent mechanisms.

Authors: Mohammed, K; Voiculescu, I

• Publication status: Accepted
• Peer review status: Peer reviewed
• Publisher: Medical Image Computing and Computer Assisted Intervention Society
• Acceptance date: 2026-08-05
• Event title: 29th International Conference of Medical Image Computing and Computer Assisted Intervention (MICCAI 2026)
• Event location: Strasbourg, France
• Event website: https://conferences.miccai.org/2026/en/
• Event start date: 2026-09-27
• Event end date: 2026-10-01
• Language: English
• Keywords: segmentation; evaluation; slice thickness; boundary metrics; uncertainty; calibration