Cancer staging decides who gets surgery, who gets chemotherapy first, and who gets palliative care. It depends heavily on whether lymph nodes near the tumor contain cancer. Clinical staging from CT regularly misses small involved nodes, most of which are under 10 mm, and those misses push patients into lower-risk groups than they belong in. Deep models could help by finding every node in a scan, but they are trained on labels that carry the same blind spot. In real hospital data, radiologists outline only the nodes they judge "clinically significant." Small or borderline nodes are left unlabeled and end up treated as background. A detector trained this way learns to ignore exactly the small nodes that matter for under-staging.
We propose to find these missing labels using a signal that 3D models throw away. A 2D model that segments each CT slice on its own, with no view of neighboring slices, is usually a weakness. Its per-slice outputs are jagged and inconsistent. That same independence makes it useful as a source of evidence. When separate slice-wise predictions, from axial, coronal and sagittal views, agree on a compact blob with the rise-peak-fall size profile of a node, the agreement is hard to explain by chance. A 3D model's smooth output is not comparable evidence, because its large receptive field makes smoothness automatic. We use this cross-slice, cross-plane agreement to decide which unlabeled detections are real nodes the annotator skipped. Those detections become training labels, and the final 3D segmenter is retrained on them.
Core Claim
In partially annotated 3D lymph node data, agreement among context-free per-slice predictions separates real unannotated nodes from spurious detections better than two alternatives. The first is the 3D teacher's own confidence. The second is loss-based per-instance trust. The advantage should be largest for nodes under 10 mm. Relabeling with this signal should raise fully-annotated test Dice and small-node recall on the LNQ 2023 benchmark above the challenge's best reported pipeline, with the pipeline itself held fixed.
The novelty is in what counts as a trustworthy label. Existing partial-label methods trust a teacher's confidence, and existing noisy-label methods trust small training loss. We trust agreement among weak, independent views of the same volume.
Where This Sits
The problem setting comes from the LNQ 2023 Challenge. It released 383 partially annotated training chest CTs and a fully annotated 100-scan test set. It documented the "false negative trap," in which standard losses penalize a model for finding unlabeled nodes. Top teams combined three fixes: 3D nnU-Net pretraining on the fully annotated NIH CT Lymph Node set, teacher pseudo-labels, and masking unannotated voxels out of the loss. The challenge reports struggles on bulky, merged nodal masses. It also says nothing about which of its pseudo-labels were correct. That gap motivates this work.
Two lines of work supply the method. LN-Tracker treats the z-axis as time and tracks nodes across slices. It gains 2.7% average sensitivity at about 17% of nnDetection's compute. That shows cross-slice persistence is a strong cue for nodeness, but LN-Tracker uses the cue only for detection against fixed ground truth, never to audit labels. Ke Yan et al. merge 2.5D detections into 3D by seeding from the most confident slice. Our per-candidate slice matching borrows that lesion-centric logic.
Other work constrains the design. Holger Roth et al. document that independent 2D predictions are discontinuous across slices. So we never use the 2D model as the final predictor, only as a witness. LNMBench shows that small-loss sample selection discards rare, hard, true cases on imbalanced medical data, and that methods tuned on synthetic flips fail on real clinical noise. Both warnings apply directly here. Small true nodes are high-loss, so loss-based trust would prune them. And our simulated noise must mimic the real "clinically significant only" rule, not random deletion. LiNC learns a trust scalar per sample and reports a noise-detection AUC of 0.9837 on 2D OrganSMNIST classification. It is the strongest general noisy-label competitor we can adapt to node instances, and its authors note it fails on systematic, guideline-driven noise. Partial annotation is noise of exactly that kind.
We borrow a principle, and set aside the method, from DiFF. DiFF defines the reference mask among conflicting 2D raters by downstream diagnostic value. We share its view that the reference standard should be chosen, not assumed. Its 2D rater-weighting network, however, has no notion of z-continuity.
The staging motivation comes from registry work. Shasha Yuan et al. and Chenzhang Ou et al. show that under-sampling nodes causes stage migration. Chaoyi Wu et al. report that 70% of metastatic nodes in their oral cancer CT set are under 10 mm. LNQ has no pathology, so we do not claim staging outcomes. We use small-node recall as the staging-relevant proxy and say so plainly.
Research Program
The pipeline has three stages.
In the first stage we build a teacher and propose candidates. We train a 3D nnU-Net teacher using the same recipe as the LNQ top teams: NIH pretraining, loss masking of unannotated voxels, and a lung exclusion mask. Every connected component the teacher predicts that does not overlap an annotation becomes a candidate unlabeled node.
In the second stage we score each candidate with independent witnesses. We train three lightweight 2D segmenters, one each on axial, coronal and sagittal slices. Each sees a single slice with no neighbors, and each uses the same partial labels and masking. For every candidate we compute three quantities. The first is the fraction of slices in its extent, in each plane, where the 2D model fires at the matching location. The second is agreement across the three planes. The third is the shape of the in-plane area profile along each axis. A node gives a single-peaked profile, while a vessel gives a long, flat, tubular profile. A small logistic model combines these into a witness score. We calibrate it on a simulation, described below, where the hidden truth is known. Using orthogonal planes matters for thick-slice scans: under 4–5 mm z-spacing, the axial witness has few slices, but the coronal and sagittal witnesses still see extent.
In the third stage we relabel and retrain. Candidates with a high witness score become positive labels, weighted by the score. Candidates with a low score become negatives, on the grounds that the annotator plausibly would have outlined them had they been nodes. Middling candidates stay masked out. A student 3D nnU-Net is then trained on the relabeled data, and we allow a single further round.
The simulation that makes this measurable uses NIH CT Lymph Node, which is fully annotated. We delete labels following a structured, size- and station-dependent rule: small nodes are removed with high probability, and nodes over a short-axis threshold are kept. This mimics "clinically significant only" rather than random deletion, answering LNMBench's sim-to-real warning. Because we know which deleted labels are real, the witness score can be scored directly as a detector of missing labels.
Baselines
The naive reference is a 3D nnU-Net trained on the partial labels with unannotated voxels masked out of the loss and no relabeling. It stands in for the purely weakly supervised setting, which LNQ reports reaches a median Dice of about 61.0%. We rerun it under our own pipeline to get a matched number.
The strongest existing method is the LNQ winning recipe from Skeleton Suns. It combines NIH pretraining, teacher pseudo-labeling with a confidence threshold, anatomical priors and loss masking, and reports a median Dice of 71.2% and an ASSD of 2.95 mm on the fully annotated test set (LNQ 2023). We reimplement the recipe with the same backbone, pretraining, epochs and test-set protocol as our method. We tune its confidence threshold on the 20 validation scans with a search budget equal to the one used to fit our witness model. Beating it means higher median Dice on the 100 test scans with a paired Wilcoxon test across scans at p < 0.05. We also require higher recall on nodes under 10 mm at a matched number of false positives per scan. A Dice gain concentrated on large nodes alone would not count.
The noisy-label competitor is an instance-level adaptation of LiNC. It gives each candidate and each annotated node a trainable trust scalar and clusters the scalars with a three-component Gaussian mixture. It runs on the same candidates and training budget.
Two ablations isolate the key ingredient. The first replaces the independent 2D witnesses with the same agreement and profile statistics computed on the 3D teacher's own output. If that works equally well, independence adds nothing. The second uses the axial witness alone, which tests whether orthogonal planes carry the method on thick-slice scans.
Why not just…?
Why not just threshold the 3D teacher's confidence, as the LNQ winners did? This is the real threat. A 3D model already uses volumetric context, so its confidence may already encode everything the witnesses do. Our argument is that confidence and correctness come apart on small, low-contrast nodes. On these, a 3D model trained to ignore unlabeled nodes should be systematically underconfident, since it was taught they are background. The independent 2D views are trained under the same bias but fail in different ways. The simulation settles the question directly. We compare the area under the ROC curve of teacher confidence against the witness score for telling hidden true nodes from spurious candidates, stratified by size. If the two are within noise, the idea is not worth its cost, and we will report that.
Why not just use per-instance trust or small-loss selection from the noisy-label literature? LNMBench shows that small-loss selection discards rare, hard true cases. Small true nodes are exactly such cases, so loss-based trust should be biased toward pruning them. LiNC's own authors note it absorbs systematic noise as clean, and a consistent "skip small nodes" rule is systematic. The instance-level LiNC baseline tests this. The telling measurement is recall on annotated small nodes after cleaning. If LiNC down-weights them and the witness score does not, the argument holds.
Why not just add more pretraining on fully annotated data? NIH pretraining is already in every arm, including the baselines, so any gain we report is on top of it. Pretraining teaches what nodes look like, but the partial-label fine-tuning still teaches the model to suppress unlabeled ones. Relabeling targets that second effect.
Why not just use slice-stacked 2.5D models or tracking as the final predictor? LN-Tracker and Ke Yan et al. improve detection, but both train against whatever labels they are given. They still inherit the missing labels. Our contribution is upstream of the predictor and should combine with either.
Proposed Experiments
The first experiment is the NIH simulation. We sweep the deletion rate and the size threshold. For each setting we report the ROC AUC of the witness score, teacher confidence, the self-coherence ablation and LiNC trust, each for recovering hidden true nodes, split at 10 mm short axis. We also report downstream student Dice and small-node recall, and we separately inspect false candidates on vessels and the esophagus, where tubular structures could fool the witnesses.
The second experiment is LNQ 2023 itself. We train on the 383 partially annotated scans, select on the 20 validation scans, and evaluate once on the 100 fully annotated test scans. We report median Dice and ASSD, the official metrics, alongside node-level recall stratified by size, false positives per scan, and free-response ROC curves.
The third is a thick-slice stress test. We resample the NIH scans to 5 mm z-spacing and repeat the first experiment, checking whether the coronal and sagittal witnesses keep the gain. We also report results separately on bulky, merged masses, which the challenge flagged as ambiguous. Our profile features assume discrete nodes and may fail on these masses.
The idea succeeds only if two things hold. The witness score must beat teacher confidence at recovering hidden small nodes in simulation, and that advantage must survive the self-coherence ablation. On LNQ, the method must beat the reimplemented winning recipe on small-node recall at matched false positives without losing median Dice. Failing the first condition falsifies the central claim. Passing it but failing the second means the recovered labels do not matter for the final model. Both outcomes would be worth publishing, given how widely confidence-based pseudo-labeling is used for 3D partial annotations.
References
- LNQ 2023 challenge: Benchmark of weakly-supervised techniques for mediastinal lymph node quantification
- From Slices to Sequences: Autoregressive Tracking Transformer for Cohesive and Consistent 3D Lymph Node Detection in CT Scans
- Anatomy-Aware Lymph Node Detection in Chest CT Using Implicit Station Stratification
- Deep convolutional networks for pancreas segmentation in CT imaging
- Benchmarking real-world medical image classification with noisy labels: Challenges, practice, and outlook
- LiNC: Lightweight Noise Correction via Adaptive Label Refinement
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Prognostic impact of examined lymph-node count for patients with esophageal cancer: development and validation prediction model
- The optimal number of examined lymph nodes for cancer specific death of intrahepatic cholangiocarcinoma: a population-based study
- Integrating features from lymph node stations for metastatic lymph node detection