Training a 3D medical segmentation model usually needs an expert to paint every voxel of every organ, which can take around 20 minutes for a single pancreas CT. Cheaper alternatives train on scribbles, a few quick strokes drawn inside and outside each structure. The best current scribble method is almost embarrassingly simple: train a standard network only on the scribbled voxels and ignore the rest. It comes within a few Dice points of full supervision. The points it loses sit at the boundary, where scribbles are absent and where organs like the pancreas fade into fat and bowel with almost no intensity edge.
This note proposes to recover those points with a geometric fact that scribbles already contain and current methods throw away. Any voxel inside an object lies no farther from the object's boundary than it does from the nearest background scribble. Any voxel outside lies no farther from the boundary than from the nearest foreground scribble. These two distances therefore give every voxel in the volume, labeled or not, a guaranteed interval for its signed distance to the true boundary. The interval is tightest exactly where a foreground and a background stroke bracket an edge. We train a network to predict a signed distance map, a per-voxel value giving distance to the boundary with negative sign inside, that stays within these intervals and behaves like a true distance function. The network runs as the fine stage of a coarse-to-fine pipeline, which works at native resolution inside a crop around a coarsely located organ, so distances are measured in real millimetres.
Core Claim
Scribbles determine more than labels at the scribbled voxels. They also determine hard upper and lower bounds on the signed distance to the boundary at every voxel. Supervising a signed-distance output with these bounds, together with a label-free regularizer that forces the output to behave like a distance, should close a substantial part of the remaining gap between scribble and dense supervision on low-contrast abdominal targets. The gains should appear mainly in boundary metrics. The same idea also changes how annotation time is best spent: a few short strokes that bracket an ambiguous edge should be worth more than long interior strokes.
Write $F$ and $B$ for the sets of foreground and background scribble voxels. Let $d_F(x)$ and $d_B(x)$ be the physical (mm) distances from voxel $x$ to the nearest voxel of each set, and let $\phi(x)$ be the predicted signed distance, negative inside the object. If $x$ is inside, the straight segment from $x$ to its nearest background scribble must cross the boundary, so $|\phi(x)|\le d_B(x)$. The symmetric argument applies outside. Together these give a bound that holds for any object shape or topology:
$$ -d_B(x) - \epsilon \;\le\; \phi(x) \;\le\; d_F(x) + \epsilon , $$
where $\epsilon$ is a small slack for strokes that stray slightly across the true edge. The ordinary scribble labels fall out as a special case. On a foreground stroke $d_F=0$, so the constraint forces $\phi\le 0$ (inside), and on a background stroke it forces $\phi\ge 0$. The training loss is
$$ \mathcal{L} = \sum_x \Big[\mathrm{relu}\big(\phi(x) - d_F(x) - \epsilon\big) + \mathrm{relu}\big(-d_B(x) - \epsilon - \phi(x)\big)\Big] + \lambda \sum_x \big(\lVert \nabla \phi(x) \rVert - 1\big)^2 + \mathcal{L}_{\mathrm{pL}} . $$
The middle term, called an eikonal penalty, pushes $\phi$ to have unit gradient like a true distance field. That is what carries tight constraints near bracketed edges out to the unbracketed parts of the surface. $\lambda$ weights this term, and $\mathcal{L}_{\mathrm{pL}}$ is the standard partial cross-entropy plus Dice on scribbled voxels, computed on probabilities obtained from $\phi$ through the smooth step $\sigma(-\phi/k)$ with temperature $k$.
Relation to Prior Work
Two papers supply the signed-distance machinery. Shape-Aware Organ Segmentation by Predicting Signed Distance Maps regresses signed distance maps and links them to probabilities through a smooth step, cutting Hausdorff distance on a hippocampus CT set from 23.57 mm to 5.40 mm. E2Net supervises an edge branch with normalized distance maps inside a two-stage liver pipeline. Both compute their target maps from fully dense masks. We keep their output representation but replace the exact target with an interval derived from scribbles, which neither paper can use. These papers motivate the choice of representation. They are not competitors under sparse labels.
Two papers set the constraint. Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation and Revisiting 3D Medical Scribble Supervision show that elaborate pseudo-labeling methods collapse off cardiac data, with ShapePU at 0.369 and CycleMix at 0.559 mean Dice on their seven-task ScribbleBench. A plain 3D nnU-Net with partial loss reaches 0.813 against 0.856 for dense labels. The lesson we take is that any added signal must never invent labels that might be wrong. Our interval constraint fits this: it is a provable consequence of correct scribbles, not a guess propagated along image edges. That distinguishes it from the propagation approach of Learning to Segment Medical Images with Scribble-Supervision Alone, whose random-walk and CRF (conditional random field, an edge-following smoother) relabeling depends on intensity gradients. Those gradients are what low-contrast boundaries lack, and the same weakness shows up in GrabCut seeding in Weakly Supervised Lesion Co-Segmentation on CT Scans.
The pipeline structure comes from coarse-to-fine work, in which a cheap model finds the organ and a second model segments a native-resolution crop around it: Towards dense volumetric pancreas segmentation and A 3D Coarse-to-Fine Framework for Volumetric Medical Image Segmentation. The latter also reports the known weakness of cascades, cases with 0% Dice when the coarse stage misses a pathological pancreas. Bounding Maps for Universal Lesion Detection shows that soft spatial grading beats hard thresholds for proposals. We borrow that intuition for crop margins but not its linear heuristic. All of these cascades are trained on dense masks. Here both stages see only scribbles.
Research Program
The coarse stage is a 3D nnU-Net trained with partial loss on downsampled volumes. It outputs a probability map, and the crop is the region where that map exceeds a low threshold, dilated by a margin that grows with the coarse stage's uncertainty. This reduces truncation of infiltrative or irregular targets. We measure truncation directly as the fraction of ground-truth foreground falling outside the crop. The fine stage is the same backbone with a single signed-distance output head, trained with the loss above on native-resolution crops. Distances are computed in millimetres using voxel spacing so that anisotropic CT is handled correctly. Multi-class tasks use one signed-distance channel per class, each one-versus-rest. Background scribbles of other classes count toward $B$.
The analysis has three parts. The first checks the bound empirically: we measure how often the dense-label signed distance violates the scribble interval on ScribbleBench. This quantifies how much slack $\epsilon$ real strokes require and whether any assumption fails. The second maps where the interval is informative by plotting interval width against distance to the true boundary. This predicts which tasks can benefit. The third is an annotation study at matched cost, using scribbles simulated from dense labels. It compares ScribbleBench-style interior strokes with "bracket" strokes, short foreground/background pairs placed across the edges the coarse stage flags as uncertain.
Baselines
The naive reference point is the coarse stage alone upsampled to full resolution. It shows what localization by itself buys and bounds the cost of the pipeline. The strongest existing method is the single-stage 3D nnU-Net with partial loss from Revisiting 3D Medical Scribble Supervision. It reports 0.813 mean Dice on ScribbleBench against a dense upper bound of 0.856. It is the strongest competitor because it has already beaten every specialized scribble method on exactly our target datasets, and because it adds nothing that could be wrong. We run it with the authors' released scribbles and training schedule, on the same splits, with the same epoch and GPU budget as our full pipeline. A beat means a higher mean Dice on the low-contrast tasks and a lower 95th-percentile Hausdorff distance, with non-overlapping bootstrap confidence intervals across cases.
A second strong baseline is the nnU-Net two-stage cascade trained with partial loss. It has the same coarse-to-fine structure as ours but no distance head, and it isolates whether the cascade alone explains any gain. The key ablation removes the interval term while keeping the signed-distance head, eikonal penalty, and cascade. It tests whether the bounds, rather than the output representation, do the work. A companion ablation drops the eikonal term to test whether interval information actually propagates. A fourth comparison applies the interval loss in a single-stage model. If it matches the cascade, the coarse-to-fine component contributes only compute savings, and we will say so. Dense-label signed-distance regression in the style of Xue et al. serves as the upper reference. For pancreas we also note the dense coarse-to-fine result on the JHMI pathological set, 80.56% mean Dice from Zhu et al., as a ceiling rather than a competitor. We found no reported scribble-supervised result on Lacuna for NIH or JHMI pancreas.
Why not just…?
Why not just use nnU-Net with partial loss and accept the 4.3-point gap? This is the real threat. The case against it is structural: partial loss gives zero gradient to every unscribbled voxel, and those voxels include all boundary voxels. The interval loss gives every voxel a correct, if loose, constraint at no additional annotation cost. Whether that matters is empirical. If the head-to-head at matched compute shows no Dice or Hausdorff improvement on AMOS, WORD, LiTS, and KiTS, the proposal fails on its main claim.
Why not just expand the scribbles into dense pseudo-masks with a CRF or GrabCut before training? ScribbleBench already shows that propagation-based methods fall far below partial loss off cardiac data. Can et al. and Agarwal et al. both tie their pseudo-label quality to intensity edges that leak at low contrast. Our constraint never assigns a label a voxel might not deserve. We still include a DenseCRF-expanded nnU-Net as a baseline to confirm this on our splits.
Why not just ask annotators to scribble near the boundaries and keep partial loss? This is a serious alternative, because boundary-placed strokes give partial loss the edge signal it lacks. The annotation study settles it by crossing stroke placement (interior versus bracket) with loss (partial versus interval) at equal simulated annotation cost. If partial loss with bracket strokes matches interval loss with bracket strokes, the contribution is an annotation protocol rather than a training objective, and we would report it that way. We expect the interval loss to win because a bracket constrains a whole neighborhood of the surface through the eikonal term, not only the stroked voxels.
Why not just use a 2D two-stage edge network like E2Net? It needs dense masks to build its edge targets. Its authors also report difficulty on small low-contrast tumors without through-plane context, which is the regime we target.
Proposed Experiments
The primary benchmark is ScribbleBench with its released scribbles. The pre-registered prediction is that gains concentrate on the low-contrast abdominal tasks (AMOS, WORD, LiTS, KiTS) and are near zero on ACDC and MSCMR. On those cardiac sets the blood-pool contrast is high and partial loss is already close to dense. A uniform gain everywhere would suggest the improvement comes from capacity rather than boundary information. Secondary experiments use NIH Pancreas and, if accessible, JHMI Pathological Pancreas, with scribbles simulated from dense labels by the ScribbleBench procedure. Metrics are Dice, 95th-percentile Hausdorff distance, and normalized surface Dice at a 2 mm tolerance, reported per case with bootstrap intervals. Per-case reporting also lets us count catastrophic failures, cases below 10% Dice, which relate directly to the 0% cascade failures of Zhu et al.
The decisive checks are three. The interval-loss ablation must show that the bounds, not the distance head, account for most of any gain. The empirical violation rate of the bound on real scribbles must be small enough that a slack of a millimetre or two suffices. The matched-cost annotation study must show whether bracket strokes plus interval loss beat bracket strokes plus partial loss. A useful target is recovering at least half of the gap between scribble and dense supervision on the low-contrast tasks while reducing 95th-percentile Hausdorff distance relative to the partial-loss nnU-Net. If the gain appears only in the single-stage variant or only with bracket strokes, the note's claim narrows accordingly, and the experiments are designed so that we can tell which of these holds.
References
- Shape-Aware Organ Segmentation by Predicting Signed Distance Maps
- E2Net: An Edge Enhanced Network for Accurate Liver and Tumor Segmentation on CT Scans
- Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation
- Revisiting 3D Medical Scribble Supervision: Benchmarking Beyond Cardiac Segmentation
- Learning to Segment Medical Images with Scribble-Supervision Alone
- Weakly Supervised Lesion Co-Segmentation on Ct Scans
- Towards dense volumetric pancreas segmentation in CT using 3D fully convolutional networks
- A 3D Coarse-to-Fine Framework for Volumetric Medical Image Segmentation
- Bounding Maps for Universal Lesion Detection