Most 3D medical scans are anisotropic: the scanner sees fine detail within each slice but leaves several millimetres of tissue unobserved between slices. The ATLAS liver MRI data, for example, has 0.68–1.4 mm resolution within a slice and 2.0–4.6 mm between slices (Lambert et al., 2023). A segmentation network should therefore be less sure about a voxel whose nearest organ edge lies in the next slice than about one whose edge is visible in the same slice. Current calibration methods treat both kinds of voxel the same way. They either add one calibration penalty averaged over every voxel or choose one global confidence threshold. Calibration here means making a stated 80% confidence right 80% of the time. We argue that this uniformity forces a bad trade. To make the poorly observed through-slice voxels honest, the model also becomes vaguer at well-observed in-plane edges, and it loses sharpness there for no reason. Sharpness is how tightly the predicted probability or mask concentrates.
We propose to split the calibration problem by the geometry of missing context. Each voxel is assigned to a group using only the model's own prediction. The groups separate voxels whose nearest predicted boundary is across a slice gap from voxels whose nearest boundary is within the slice, and they are binned by distance measured in units of slice thickness. Calibration is then trained within each group, and statistical error guarantees are fitted with group-specific thresholds. The hypothesis can be tested directly. In anisotropic data, calibration error should concentrate in the through-slice groups and grow with slice thickness. Treating those groups separately should then give masks that are equally safe but tighter than masks from uniform calibration.
Core Claim
In anisotropic volumes there is no single calibration–sharpness frontier. There is one frontier per geometric context, and the through-slice frontier is markedly worse. Methods that pool all voxels sit on a compromise that is too confident through-slice and too diffuse in-plane. A calibration objective and a conformal threshold family that are conditioned on predicted through-slice versus in-plane boundary geometry should do two things relative to pooled methods. First, they should lower calibration error inside the through-slice group. Second, at a fixed guaranteed false-negative rate, they should produce more compact masks. Conformal prediction is a post-hoc procedure that turns held-out data into finite-sample error guarantees. The effect should shrink to zero as data become isotropic, and that dose-response is the prediction that makes the claim falsifiable.
Relation to Prior Work
The motivating evidence comes from three places. Dheur and Ben Taieb (2023) showed across 57 tabular datasets that post-hoc recalibration achieves calibration mostly by widening predictive distributions, while calibration penalties applied during training preserve sharpness. Chung et al. (2021) showed that standard losses can trade honesty for sharpness invisibly. Neither studied spatial structure. Barfoot et al. (2025) brought calibration-in-training to volumetric segmentation with a differentiable binned calibration loss (sL1-ACE). They observed that this loss slightly blurs probability boundaries, which is the trade-off we want to localise, but they bin all voxels together without regard to slice geometry. Liang et al. (2025) tie confidence to signed distance from the ground-truth boundary, which is the closest conceptual neighbour. Their formulation is 2D, assumes a fixed logistic distance-to-probability shape, and uses ground-truth distance, which is unavailable at test time. We stratify by predicted geometry instead, so the same groups exist at inference and can carry conformal guarantees, and we distinguish the direction of the missing context rather than only its distance.
Two lines of work supply method components. Tan et al. (2025) (Conformal Lesion Segmentation, CLS) select one global threshold across held-out volumes so that the per-volume false-negative rate (FNR) stays below a target with high probability. They measure the resulting over-segmentation with a prediction-compactness score. We reuse their calibration procedure but replace the single threshold with a one-parameter family of group-specific thresholds. COMPASS shows that conformal intervals become much tighter when calibration exploits model structure rather than treating outputs as a black box. Its experiments are on 2D images only, and it does not ask what anisotropy does.
Other work constrains the design. Lambert et al. (2023) show that an anisotropy-aware hybrid architecture outperforms an isotropic 3D U-Net on ATLAS, and that uncertainty correlates strongly with small lesion size (ρ = −0.82). Better architecture therefore helps, and any gain we report must hold on top of it. Lambert et al. (2024) handle covariate shift by spatially averaging features, which by design cannot see slice-local effects. We do not tackle shift, and exchangeability between calibration and test volumes remains an assumption, as in CLS.
Research Program
The first component is a diagnostic. For each voxel we take the network's predicted mask and compute two distances in millimetres to its boundary: the 3D distance $d_3$ and the in-slice distance $d_2$, which ignores neighbouring slices. When $d_3 \ll d_2$, the nearest edge lies above or below, in the poorly observed direction. Voxels are binned by $d_3$ expressed in slice thicknesses and by the ratio $d_3/d_2$, which gives a small set of groups $s \in \{1,\dots,S\}$ with $S$ around 6–9. We report calibration error per group, as well as the pooled figure, before proposing any fix. If through-slice groups are not clearly worse, the project stops there with a useful negative result.
The second component is a training objective, a group-conditioned version of the sL1-ACE loss:
$$ \mathcal{L} = \mathcal{L}_{\text{seg}} + \lambda \sum_{s=1}^{S} w_s \, \text{ACE}_s , $$
Here $\mathcal{L}_{\text{seg}}$ is the usual Dice plus cross-entropy loss, and $\text{ACE}_s$ is the soft-binned average calibration error computed only over voxels in group $s$. The weight $w_s$ is proportional to the square root of the group's voxel count, so that small through-slice groups are not drowned out, and $\lambda$ is the penalty weight. Groups are recomputed from the current prediction every few hundred iterations. Distance transforms on downsampled masks keep this cheap.
The third component is a conformal step. Each group gets a threshold $t_\theta(s) = b_s + \theta$. The offsets $b_s$ are fitted on one held-out split to maximise compactness at roughly the target FNR. The scalar $\theta$ is then chosen on a second split using exactly the CLS quantile rule. FNR is monotone in $\theta$ and the offsets are fixed before $\theta$ is chosen, so the finite-sample FNR guarantee of CLS carries over unchanged. Any compactness gain comes only from spending the error budget unevenly across geometric contexts.
Baselines
The naive reference is an nnU-Net-style 3D network trained with Dice plus cross-entropy and thresholded at 0.5. On KiTS21, Tan et al. (2025) report an FNR of about 0.53 for this default across their backbones. It anchors how badly uncontrolled models under-segment.
The strongest existing training-time method is pooled sL1-ACE (Barfoot et al., 2025), which the authors report reduces Macro-ACE by 33% and Micro-ACE by 51% on average over ACDC, AMOS, BraTS and KiTS. We will rerun it on our splits with the same backbone, the same epochs, and the same grid of $\lambda$ values. Our method beats it if it reduces through-slice group calibration error by a margin outside seed variance at equal or better Dice and equal pooled calibration error.
The strongest existing guarantee method is CLS with a global threshold. On KiTS21 it brings FNR to about 0.10. We found no reported prediction-compactness number on Lacuna for CLS on KiTS21 or ULS23, so we will reproduce it. Beating it means equal empirical FNR, which both methods guarantee, with significantly higher compactness, using the same calibration and test splits and the same frozen base model.
The most dangerous baseline is post-hoc temperature scaling fitted separately per geometric group. Temperature scaling divides the network's output scores by one learned constant, and here we learn one constant per group on the same calibration data. If this matches our method, then the training loss is unnecessary and only the grouping matters.
Two ablations isolate the key ingredient. In the first, we replace geometric groups with random groups of matched sizes. In the second, we use distance-only groups that ignore the in-plane versus through-slice ratio. This tests whether direction matters, as opposed to merely conditioning on more groups. On ATLAS we also run every method on top of AHUNet, so gains are not artefacts of a weak isotropic backbone. For reference, AHUNet reaches liver Dice 0.94 and tumour Dice 0.57 there.
Why not just…?
Why not just use one global temperature scaling? It multiplies every voxel's scores by the same number, so it can only move the pooled calibration curve. It cannot be honest through-slice and sharp in-plane at the same time, and Barfoot et al. (2025) found it weaker than training-time losses on volumetric data. It is included as a baseline. If our diagnostic shows no group disparity, global scaling will be as good as anything, and we will report that.
Why not just temperature-scale per geometric group after training? This is the real threat, and it uses our own grouping. Post-hoc rescaling can make each group calibrated, but on the evidence of Dheur and Ben Taieb (2023) it should get there by flattening probabilities, not by improving the ranking of voxels. The decisive comparison is therefore within-group compactness at fixed FNR and within-group discrimination (AUROC of voxel foreground versus background), not calibration error alone. If per-group scaling ties on these, we will recommend it as the practical method and keep only the diagnostic and the conformal threshold family as contributions.
Why not just resample to isotropic voxels or use an anisotropy-aware architecture? Interpolation cannot create anatomy that was never imaged, so through-slice ambiguity remains. Architecture changes are complementary rather than competing: AHUNet improves accuracy (Lambert et al., 2023), yet Tan et al. (2025) found strong 3D backbones still miss about half of lesion volume at default thresholds. Running our method on AHUNet and on resampled inputs tests whether the gap survives.
Why not just smooth the probability map in 3D? Roth et al. (2015) raised pancreas Dice from 56% to 68% with 3D Gaussian smoothing of 2D predictions. Smoothing blurs evenly in all directions, however, which is exactly the uniform sharpness loss we hypothesise. We include anisotropic Gaussian smoothing followed by CLS as a baseline and expect lower compactness at matched FNR.
Proposed Experiments
The data are KiTS21 and KiTS23 kidney CT, AMOS multi-organ CT, ATLAS liver MRI, ACDC cardiac cine-MRI and the ULS23 lesion sets used by CLS. Slice spacing will be read from the image headers, and BraTS, the least anisotropic of the set, serves as a control. The backbones are nnU-Net, Swin-UNETR and, on ATLAS, AHUNet, each with three seeds. The metrics are Dice and surface Dice, for context against the KiTS21 top Dice of 0.908 reported by Heller et al. (2023). We also report pooled and per-group calibration error, empirical FNR against the target ε, CLS prediction compactness, and calibration on the first and last slices of each structure.
The first decisive check is a dose-response experiment. We take near-isotropic CT volumes, drop slices to simulate thicknesses of 1×, 2×, 3× and 5× the original, and retrain. The prediction is that the through-slice calibration gap of pooled methods, and our compactness advantage, both grow with thickness and vanish at 1×. A flat curve would refute the core claim.
The second check is the comparison with per-group temperature scaling at matched calibration. Our training loss must win on within-group discrimination and compactness, not only on calibration error.
The third check uses the random-group ablation. If random groups recover most of the gain, the gain comes from extra flexibility rather than from geometry, and the hypothesis fails.
All conformal results report empirical FNR across 100 random calibration/test splits to confirm that the group-wise threshold family keeps the CLS guarantee.
References
- Anisotropic Hybrid Networks for Liver Tumor Segmentation with Uncertainty Quantification
- A Large-Scale Study of Probabilistic Calibration in Neural Network Regression
- Beyond Pinball Loss: Quantile Methods for Calibrated Uncertainty Quantification
- Average Calibration Losses for Reliable Uncertainty in Medical Image Segmentation
- We Care Each Pixel: Calibrating on Medical Segmentation Model
- Conformal Lesion Segmentation for 3D Medical Images
- COMPASS: Robust Feature Conformal Prediction for Medical Segmentation Metrics
- Robust Conformal Volume Estimation in 3D Medical Images
- Deep convolutional networks for pancreas segmentation in CT imaging
- The KiTS21 Challenge: Automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary-phase CT