When doctors label medical scans with quick scribbles instead of full outlines, the scribbles almost always sit safely inside the organ or tumor. That leaves the network with no training signal at the boundary itself, which is where low-contrast targets are hardest. This proposal uses the scribbles as more than a loss target. In each image, the network pools its features separately over the scribbled inside and the scribbled outside. From that it measures, for every feature channel, how well that channel tells target from surroundings in this particular scan. It then boosts those channels before the final full-resolution layers that draw the boundary. At test time no scribbles exist, so the same "anchors" come from the network's own confident interior and exterior. The intended result is a per-image contrast lens that needs no boundary annotation and no external edge detector.
The idea joins two familiar parts with a change to how one of them pools. Squeeze-and-excitation (SE) channel gating is the familiar part: a small module that averages each channel over the whole image and turns that average into a 0–1 weight per channel. The other part is a dilated convolutional network that never downsamples. Dilated convolutions spread their kernel taps apart, so the receptive field grows without shrinking the feature map. The change is to replace SE's whole-image average with a contrast measured between anchored regions. The test is deliberately harsh. The method must beat the plain nnU-Net with partial loss, which Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation shows already sits within about 4 Dice points of full supervision across seven diverse tasks. It must also beat it on the low-contrast tumors and organs where that remaining gap should concentrate.
Core Claim
The remaining gap between scribble and dense supervision on diverse 3D benchmarks is mostly a boundary-band gap on low-contrast targets. It comes from two sources. The first is that no gradient reaches the target–background interface. The second is that standard channel gating cannot find the few channels that carry separating evidence in a given scan, because a whole-image average is dominated by background. A spatially anchored squeeze addresses both, since it computes channel weights from target-versus-surroundings contrast rather than from global means. Placed in a downsampling-free dilated network, it should close a measurable fraction of that gap. The closed fraction should grow as target contrast falls, and it should do so with one fixed hyperparameter setting across all tasks.
Each input direction plays a different role. Channel recalibration is the mechanism being redesigned. Resolution-preserving dilation is the host architecture, chosen because it has no upsampling path that interpolates across unlabeled boundary bands. Low-contrast boundary ambiguity is the regime where the effect should appear. The annotation bottleneck is the constraint that rules out boundary supervision entirely.
Why This Is Not Already Done
The closest prior systems each lack a different piece.
SUNet++ places SE layers in UNet++ skip connections to keep small liver nodules from washing out, reaching LiTS tumor Dice 0.4724 against nnUNet's 0.4501. It uses ordinary global-average squeezes, dense labels, and a downsampling backbone. Its reported sensitivity to the reduction ratio is a warning this proposal takes seriously.
TDNet does combine dilation with scribbles. It uses three decoders with fixed dilation rates (1, 3, 6) and an uncertainty-weighted pseudo-label consistency loss, reaching 85.74% DSC and 7.84 mm HD95 on WORD. Its channels are never reweighted per image, and it has only been shown on abdominal organs with per-slice scribbles.
Boundary Perception Guidance recovers boundaries under scribbles by regressing toward an off-the-shelf edge detector. On PASCAL VOC it reaches 99.5% of full supervision. Its edge detector is built for step edges in natural photographs, and those are exactly what soft-tissue CT and MRI lack.
Learning to Segment Medical Images with Scribble-Supervision Alone pulls labels toward intensity edges with a dense conditional random field (CRF). Its authors report strong CRF hyperparameter sensitivity, and intensity edges are weakest precisely where this proposal aims.
Boundary-attention designs such as MISNet and the reverse-attention network of Parallel Res2Net motivate the focus on perimeters, but both need dense boundary masks. None of these pools channel statistics over label-defined regions to choose which channels to trust at the boundary of a specific image.
Two papers constrain the design.
MeshNet, in State-of-the-Art Stroke Lesion Segmentation at 1/1000th of Parameters, supplies the host architecture. It is a 10-layer symmetric dilation network on full-resolution volumes that reaches 0.876 Dice on chronic stroke with about 147k parameters. Its authors also report that it under-segments thin extensions and irregular borders, which is the failure we target.
Revisiting 3D Medical Scribble Supervision supplies the main negative result. Specialized scribble methods tuned on cardiac MRI collapse on ScribbleBench. ShapePU falls to a mean Dice of 0.369, and CycleMix to 0.559. Any new module therefore has to be evaluated on all seven tasks with frozen hyperparameters, or it proves nothing.
Research Program
The host is a MeshNet-style network. It has an expansion stage with dilations rising from 1 to 16 and a refinement stage with dilations falling back to 1. The expansion stage ends in a lightweight auxiliary segmentation head. Between the two stages sits the anchored recalibration module.
For an image with feature map $X$, let $F$ be a set of target "anchor" voxels and $B$ a set of background anchor voxels. During training these are the scribbled voxels. The module computes, for each channel $c$, a contrast score
$$z_c = \frac{\mu_F(X_c) - \mu_B(X_c)}{\sqrt{\tfrac{1}{2}\left(\sigma_F^2(X_c) + \sigma_B^2(X_c)\right)} + \epsilon},$$
where $\mu_F$ and $\sigma_F$ are the mean and standard deviation of channel $c$ over $F$, $\mu_B$ and $\sigma_B$ are the same over $B$, and $\epsilon$ is a small constant. The vector of scores is concatenated with the ordinary global average. It then passes through the usual two-layer bottleneck and a sigmoid to give per-channel weights. These weights scale $X$ before the refinement layers. For multi-class tasks, one score per class versus all others is computed and the results are stacked.
The training-time anchors are only half the design, because the module must also run when no scribbles exist. At inference, $F$ and $B$ come from the auxiliary head's predictions above a high confidence threshold, eroded by a few voxels. To keep training and test conditions aligned, training mixes scribble anchors with self-predicted anchors on a schedule. That keeps the module from depending on anchor sources it will never see at test time.
If the auxiliary head finds no confident foreground, the module falls back to the plain global average. Small missed tumors would trigger this fallback, so we report its frequency and lesion-wise recall separately.
The loss is the partial cross-entropy on scribbled voxels, identical to the nnU-Net with partial loss baseline. Holding the loss fixed means any gain is attributable to architecture and recalibration rather than to extra supervision tricks.
All hyperparameters are set once on a held-out split of a single non-cardiac ScribbleBench task, then frozen for every other task. These include the reduction ratio, the confidence threshold, the erosion radius and the anchor mixing schedule. CT volumes that do not fit in memory at native resolution are processed in large overlapping patches. A scan's anchors are pooled over the patch, and patch size is reported as a variable, because the earlier thoracic work 3D Enhanced Multi-scale Network shows patching trades away global context.
The analysis has three parts. First, we decompose the scribble-to-dense gap of every model into a boundary band and an interior. The boundary band covers voxels within $k$ mm of the true contour, for $k$ of 2 and 5. Second, we measure per-case target contrast as the intensity separation between the ground-truth target and a surrounding ring, and stratify all results by it. Third, we check whether the channels the module upweights are in fact more separating inside the boundary band than globally averaged channels are. That would show the mechanism works for the stated reason rather than as extra capacity.
Baselines
The naive reference point is nnU-Net trained with partial loss on ScribbleBench scribbles. It is also, uncomfortably, the strongest existing method. Embarrassingly Simple Scribble Supervision and Revisiting 3D Medical Scribble Supervision report it at 0.813 mean Dice across the benchmark, against a dense-supervision upper bound of 0.856. Specialized methods sit far below it. We run it with its own self-configuration on identical splits and scribbles, and match total GPU-hours by giving it extra epochs if needed. Beating it means two things. The mean Dice must not fall below 0.813 when reproduced. The boundary-band Dice on the low-contrast subset must close at least a quarter of that subset's measured gap to dense supervision, holding across seeds. A gain on overall Dice alone would not count.
The second baseline is the dense-supervision upper bound, nnU-Net trained on full masks. It defines the gap whose closure we measure.
The strongest dilation-plus-scribble competitor is TDNet, which reports 85.74% DSC and 7.84 mm HD95 on WORD. We rerun it on all seven tasks with the frozen-hyperparameter protocol. On WORD we compare against both its published numbers and our rerun.
The strongest boundary-recovery competitor is the scribble pipeline of Can et al., run with its dense CRF and uncertainty gating on the nnU-Net backbone. Its CRF parameters are tuned on the same single development task as ours, then frozen.
The ablations isolate the key ingredient. The first is the same MeshNet host with no recalibration. The second is the host with standard global-average SE at the same location and parameter count. The third is the host with anchored recalibration but training-time scribble anchors only. The fourth grafts anchored recalibration into the nnU-Net decoder. The second ablation is decisive: if plain SE matches anchored SE, the contrast squeeze is not the reason for any gain. The fourth determines whether the resolution-preserving host matters at all.
Why not just…?
Why not just use nnU-Net with partial loss and stop? This is the real threat, and we do not argue around it. The evidence that it leaves room is indirect. Its partial loss gives no gradient at unlabeled boundary voxels. Its decoder upsamples coarse features across exactly those bands. The 4-point gap to dense training is measured but has never been located. The boundary-band decomposition will show where the gap actually lives. If most of it is interior, or concentrated in high-contrast targets, the motivation for this proposal fails and we will report that.
Why not just add ordinary squeeze-and-excitation blocks? Global averaging over a scan where a liver tumor occupies well under one percent of voxels mostly measures background. The SUNet++ authors also found such blocks sensitive to the reduction ratio across tasks. Whether this matters is settled directly by the matched-parameter SE ablation and by the channel-separability analysis. We will accept the outcome if plain SE does equally well.
Why not just post-process with a CRF or threshold polarization? CRFs follow intensity edges, which are weakest exactly at low-contrast interfaces, and Can et al. report strong parameter sensitivity. The size-based polarization of Parallel Res2Net assumes a known object scale, which LiTS tumors violate. We include the CRF variant as a full baseline. Our contrast stratification predicts its gains shrink as contrast falls while ours should not. If both follow the same curve, our mechanism adds nothing over intensity edges.
Why not just use a transformer for global context, as ScribFormer does? ScribFormer reaches 0.888 Dice on cardiac MRI. However, the ScribbleBench study reports that it does not carry over beyond cardiac data. Global context is also not the missing ingredient we hypothesize; boundary-channel selection is. We therefore spend the compute budget on nnU-Net and TDNet comparisons rather than on transformer variants.
Proposed Experiments
The primary experiment trains all models on all seven ScribbleBench tasks, using the benchmark's interior and border scribbles. Hyperparameters are frozen after tuning on one development task, and three seeds are run. We report Dice, 95th-percentile Hausdorff distance, and boundary-band Dice at 2 and 5 mm. The headline quantity is gap closure, the fraction of the dense-minus-partial-loss gap recovered, computed per task. The pre-specified low-contrast subset is the kidney tumors in KiTS2023, the liver tumors in LiTS, the BraTS2020 tumor subregions, and the pancreas and gallbladder in AMOS2022 and WORD.
A second experiment probes MeshNet's own reported failure mode on the ARC stroke cohort. MeshNet reports 0.876 Dice there under dense labels. We generate scribbles with the same procedure as ScribbleBench and measure whether anchored recalibration specifically recovers thin lesion extensions. These are scored as recall over the ground-truth skeleton beyond a distance from the lesion core.
The decisive checks are three. First, gap closure must increase as target contrast decreases; uniform gains would point to extra capacity, not contrast selection. Second, anchored SE must beat matched global SE on boundary-band Dice in the low-contrast subset. Third, the test-time self-anchored version must keep most of the gain of an oracle version that receives scribble anchors at inference; otherwise the mechanism is not usable without a human in the loop.
Three outcomes would falsify the proposal. Failing to beat nnU-Net with partial loss on boundary-band metrics would be one. A mean Dice regression below its 0.813 would be another. Equal results from the grafted nnU-Net variant would be a third, but that outcome would narrow the claim to the anchored squeeze alone, which would still be worth reporting.
References
- Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation
- SUNet++: A Deep Network with Channel Attention for Small-Scale Object Segmentation on 3D Medical Images
- Scribble-Based 3D Multiple Abdominal Organ Segmentation via Triple-Branch Multi-Dilated Network with Pixel- and Class-Wise Consistency
- Boundary Perception Guidance: A Scribble-Supervised Semantic Segmentation Approach
- Learning to Segment Medical Images with Scribble-Supervision Alone
- Multi-scale information sharing and selection network with boundary attention for polyp segmentation
- Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation
- State-of-the-Art Stroke Lesion Segmentation at 1/1000th of Parameters
- Revisiting 3D Medical Scribble Supervision: Benchmarking Beyond Cardiac Segmentation
- 3D Enhanced Multi-scale Network for Thoracic Organs Segmentation
- ScribFormer: Transformer Makes CNN Work Better for Scribble-based Medical Image Segmentation