When only a few scans in a medical dataset are labeled, the usual approach is semi-supervised learning. A "teacher" network labels the unlabeled scans, and a "student" trains on those guesses, called pseudo-labels. These guesses fail most often at organ and lesion borders. That is also where the image gives the least help: the pancreas fades into neighboring fat and the liver into a tumor with almost no change in intensity. Each round of self-training then reinforces errors along the border. A classical tool for keeping borders sharp exists. The bilateral filter averages each pixel only with neighbors that are nearby and look similar, so smoothing stops at edges. The fast bilateral solver turns this into a quick, differentiable optimization. On paper it could clean up pseudo-labels: spread the teacher's confident interior predictions outward, but never across an edge.
There is a catch, and it shapes this proposal. A bilateral filter decides "look similar" from raw intensity. At a low-contrast boundary, the two sides look the same, so the filter smooths straight across the border it was meant to protect. We propose to keep the solver but change what it looks at. The solver uses a small guide image to decide which pixels count as similar. We replace raw intensity with a few channels of learned features, trained only on the labeled scans, so that pixels on opposite sides of a true anatomical border look different to the solver even when they look alike to the scanner. The guide is trained through the differentiable solver with a label-propagation task. We hide the labels in a band around each true border on labeled scans, let the solver fill them in from the interior, and penalize mistakes. The result is an edge-preserving pseudo-label refiner whose notion of "edge" is learned rather than photometric.
Core Claim
Under scarce labels, a bilateral solver whose guide image is learned for label propagation will improve boundary accuracy (95th-percentile Hausdorff distance and boundary-band Dice) of semi-supervised segmentation on low-contrast targets. It should beat both uncertainty-weighted consistency methods and the same solver guided by raw intensity. The gain should come specifically from low-contrast boundary segments, where intensity guidance is expected to hurt.
Relation to Prior Work
The method comes from The Fast Bilateral Solver. That work casts edge-aware smoothing as a least-squares problem on a downsampled "bilateral grid" of position plus reference-image color, solves it with conjugate gradients, and supplies analytic gradients. It was validated on natural-image tasks such as depth upsampling and semantic segmentation refinement, with a color image as the guide. We reuse the solver unchanged. Our contributions are a learned guide, a training signal for that guide, and placement inside a semi-supervised loop.
Two classical medical papers motivate the approach by showing that bilateral weighting matters under noise. BWFCM puts a spatial-times-intensity kernel inside fuzzy clustering for brain MRI and holds a 92.00% similarity index under 9% Rician noise, against 83.73% for plain clustering. PDB-SLIC uses bilateral filtering with histogram equalization as offline preprocessing for vertebra superpixels. Both use fixed, hand-tuned bandwidths on raw intensity. They also constrain our design, because they work where boundaries have contrast (tissue classes, bone cortex). That is the exact regime the research file's analysis of low-contrast targets says breaks down for pancreas or infiltrative tumors.
The semi-supervised methods are our competitors and show where the gap is. DyCON scales consistency penalties by per-voxel entropy instead of discarding uncertain voxels, and adds a focal contrastive loss. URPC enforces consistency across decoder scales, weighted by uncertainty. MisMatch enforces consistency between dilated and eroded feature maps. All three decide how much to trust a boundary voxel. None of them say what the label should be using image geometry, and that second decision is what the solver supplies.
Two papers use "bilateral" only as a metaphor. BARL aligns two co-trained networks in both label and representation space. Graph-BAS3Net passes graph messages between a mask decoder and a boundary decoder. Neither contains an edge-preserving filter.
The closest work in spirit is AGENet. It builds edge-aware geodesic distance fields with Fast Marching and a Sobel-based speed function to concentrate few-shot prototypes near weak borders. We differ in three ways: our edges are learned rather than Sobel, our setting is semi-supervised self-training rather than episodic few-shot matching, and our output is a refined label rather than a prototype weighting. AmbiSSL models disagreement between experts on boundaries, but generates diverse boundaries by random decoder pruning with no image constraint. Our solver could act as the image constraint that pruning lacks. We treat that as an extension, not the core claim.
Mechanism
Let the teacher's foreground probability at voxel $i$ be $t_i$, with confidence $c_i = 1 - H(t_i)$, where $H$ is binary entropy normalized to $[0,1]$. Let a small guide network $g$ map the image to a $d$-dimensional feature $f_i$, with $d = 2$ or $3$, at voxel position $p_i$. The refined pseudo-label $x$ solves
$$\min_x \; \frac{\lambda}{2}\sum_{i,j} W_{ij}(x_i - x_j)^2 + \sum_i c_i (x_i - t_i)^2, \qquad W_{ij} = \exp\!\left(-\frac{\|p_i - p_j\|^2}{2\sigma_s^2} - \frac{\|f_i - f_j\|^2}{2\sigma_r^2}\right),$$
where $\lambda$ trades smoothness against fidelity to the teacher, and $\sigma_s, \sigma_r$ are the spatial and feature bandwidths. Confident interior voxels anchor the solution. Uncertain border voxels are filled in from neighbors that the guide places on the same side. The guide gets only one training signal. On a labeled scan, we set $c_i = 0$ in a band of width $b$ around the true border, set $t$ equal to the true label elsewhere, solve, and apply cross-entropy against the true label inside the band. Gradients flow through the solver into $g$, $\sigma_r$, and $\lambda$. This makes the guide learn exactly the property the solver needs, that the two sides of a real border are far apart in feature space. It never sees pseudo-labels, so it cannot inherit the teacher's self-confirming errors.
The low guide dimension is a deliberate cost constraint. Volumetric position plus three feature channels makes a 6-dimensional grid. The research file flags high-dimensional bilateral filtering as a memory problem. We control the cost by solving only inside a region of interest: a dilated band around the teacher's predicted boundary, plus a coarse grid. This is the one place the region-of-interest direction enters the core method.
Research Program
We first build the solver layer for 3D volumes with a band-restricted region of interest. We measure its wall-clock and memory cost against a single student forward and backward pass, and keep $d$ at the largest value whose overhead stays under a fixed fraction of training time. We then train the guide by band-masked propagation on the labeled subset alone. As a sanity check before any semi-supervised training, we run the propagation task on held-out labeled scans: band Dice of solver-filled labels using the learned guide versus an intensity guide versus no smoothing.
Next we insert the refiner into a mean-teacher loop on a V-Net backbone matching DyCON's setup. The student trains on refined pseudo-labels, and the guide is refreshed periodically on labeled data. We also test whether the refiner composes with DyCON's entropy-weighted consistency rather than replacing it, since the two address different questions.
The analysis centers on local contrast. For each test boundary point, we compute the intensity difference across the ground-truth border within a fixed window. We bin points into contrast quantiles and report boundary error per bin for every method. This turns the low-contrast weakness of bilateral filtering, identified in the research file's analysis, from an assumption into a measured curve.
Baselines
The naive reference point is supervised training on the labeled fraction alone, using the same V-Net, augmentation, and schedule. It shows how much any semi-supervised method adds. Above it sits a plain mean-teacher with unrefined pseudo-labels, the direct parent of our method.
The strongest existing competitor is DyCON. It is the best semi-supervised method in the file on our target benchmarks and is designed specifically for uncertain boundaries. It reports 91.18% Dice on Left Atrium with 5% labels, 84.81% Dice on Pancreas-NIH with 20% labels, and 61.48% Dice on ISLES-2022 with 5% labels. We will rerun it from the authors' settings on our exact splits, matching the published numbers within noise before comparing. Beating it means a statistically significant HD95 reduction across seeds on Pancreas-NIH and ISLES-2022 with Dice no worse. A Dice gain alone would not count, since global Dice is known to hide boundary failures. For liver CT, Graph-BAS3Net reports 93.19% Dice on LiTS with 10% labels. It serves as a secondary boundary-aware reference, run in its native 2.5D form.
The ablation that isolates our key ingredient is the identical pipeline with the guide set to raw intensity (with a CLAHE-normalized channel for CT), with $\sigma_r$ and $\lambda$ tuned on validation. Two further ablations separate learned from learned for propagation. The first uses the teacher's own penultimate features, reduced to $d$ channels, as the guide. The second trains the guide with a generic supervised segmentation loss rather than the band-masked propagation loss.
Every method gets the same backbone, labeled and unlabeled splits, number of training iterations, and an equal budget of hyperparameter trials on the same validation set. Solver overhead is reported, and a compute-matched variant of DyCON with extra iterations is run so that our method does not win simply by using more compute.
Why not just…?
Why not just run the fast bilateral solver, or a dense CRF-style refiner, on raw intensity? This is the most serious threat, and it is the intensity-guide ablation above. Our argument is that at near-zero-contrast borders the intensity kernel cannot separate the two sides. Evidence for bilateral gains, such as BWFCM, comes from tissues with real contrast. If the intensity guide matches the learned guide in the lowest-contrast quantile bins on Pancreas-NIH, the learned guide is unnecessary, and we would report the simpler method as the finding.
Why not just rely on uncertainty weighting as DyCON does? Entropy weighting decides how loudly an uncertain voxel speaks, but it cannot correct the label at that voxel. If the teacher is confidently wrong along a border, as the file's discussion of confirmation bias on fuzzy edges describes, weighting does nothing. We test whether the refiner adds to DyCON (the combined run) rather than only competing with it. If the combination equals DyCON alone, the refiner is redundant.
Why not just add an edge loss or a boundary branch? Contour branches like MISNet work well with full supervision on polyps. The research file notes, however, that auxiliary edge decoders produce fragmented edges under heavy noise, and Sobel-based edges in AGENet are sensitive to noise and parameters. We include a Sobel edge-loss variant of the mean-teacher at matched budget. If it closes the HD95 gap on ultrasound-free CT and MRI benchmarks, our extra machinery is not justified.
Why not just learn better features with pre-training? Positional contrastive pre-training, as in PCL, is complementary rather than competing. If budget allows, we would initialize all methods from the same pre-trained encoder to check that the gain survives stronger features.
Proposed Experiments
The main benchmarks are Pancreas-NIH at 10% and 20% labels and ISLES-2022 at 5% and 10%, chosen for low contrast and irregular borders. Left Atrium at 5% serves as a higher-contrast control where we expect little gain. LiTS at 10% covers liver. Metrics are Dice, HD95, boundary-band Dice at 2-voxel and 5-voxel widths, and the per-contrast-bin boundary error curve, over three seeds with paired significance tests. A noise-robustness side experiment adds synthetic Rician noise to MRI inputs, testing whether the learned guide degrades more gracefully than the intensity guide.
The decisive check is the contrast-binned curve. The hypothesis survives only if three things hold. The intensity guide should help on high-contrast boundary segments and hurt or be neutral on the lowest-contrast bins. The learned propagation guide should improve on DyCON in those lowest bins. And the generic-feature guide should fall between them. A flat curve for the learned guide relative to intensity, or gains confined to high-contrast segments, would falsify the core claim.
A second check is the propagation sanity test on held-out labeled scans. If band-filled labels from the learned guide are not clearly better than intensity-filled ones before any semi-supervised training, we stop, because the downstream gains would have no mechanism behind them.
References
- The Fast Bilateral Solver
- Segmentation of Brain Tissues from MRI using Bilateral Filter Based Fuzzy C-Means Clustering
- Sagittal balance parameters measurement on cervical spine MR images based on superpixel segmentation
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation
- Efficient Semi-supervised Gross Target Volume of Nasopharyngeal Carcinoma Segmentation via Uncertainty Rectified Pyramid Consistency
- MisMatch: Calibrated Segmentation via Consistency on Differential Morphological Feature Perturbations With Limited Labels
- BARL: Bilateral Alignment in Representation and Label Spaces for Semi-Supervised Volumetric Medical Image Segmentation
- Graph-BAS3Net: Boundary-Aware Semi-Supervised Segmentation Network with Bilateral Graph Convolution
- AGENet: Adaptive Edge-aware Geodesic Distance Learning for Few-Shot Medical Image Segmentation
- Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation
- Multi-scale information sharing and selection network with boundary attention for polyp segmentation
- Positional Contrastive Learning for Volumetric Medical Image Segmentation