When a segmentation model outlines a small brain metastasis, it can be wrong in two different ways. It can put the edge in the wrong place because the lesion fades into surrounding brain. Or it can outline the wrong thing: a blood vessel that looks like a lesion, two lesions merged into one, or a treated lesion whose dead (necrotic) tissue looks like regrowing tumor. The probability map that standard models output mixes these two kinds of doubt into one number per voxel. We propose a different output. The network predicts, at every voxel, a small arrow pointing toward the nearest lesion edge and another pointing toward the center of the lesion the voxel belongs to. Two properties of these arrows can then be read as separate doubts. Where the edge location is uncertain, the arrows shrink. Where two candidate structures compete for the same voxels, neighboring arrows point in clashing directions. We test whether the second signal, directional clash, flags the lesion-level errors that matter clinically, and whether it keeps working when the input shifts from curated pre-treatment scans to the messier post-treatment scans a deployed system would actually see.
The stakes come from how clinical AI succeeds in practice. The one large prospective deployment in the file did not replace readers. It routed a small fraction of cases to human arbitration and added only 4–6% reading workload (Prospective implementation of AI-assisted screen reading). A segmentation model is useful in the same way only if it can say which lesions to check, and only if that signal survives the shift in case mix that separates retrospective benchmarks from live use.
Core Claim
A network trained with a squared-error loss to regress unnormalized direction vectors learns the average of the target directions it has seen for a given image appearance. The boundary head predicts an arrow toward the nearest edge. Consider a point at signed position $s$ along the edge normal, and let $F$ be the distribution of plausible edge positions given the image. The expected target is $+1$ when the edge lies beyond $s$ and $-1$ when it lies before it, so the predicted magnitude is
$$ m(s) = \left|1 - 2F(s)\right| . $$
The width of the band where $m$ is small therefore gives the spread of plausible edge positions in millimetres. We call this the magnitude signal. We do not claim it is new information. A well-calibrated probability map carries the same one-dimensional profile, since $P(\text{inside at } s) = 1 - F(s)$. We include it as a control and as a convenient readout in physical units.
The claim we do make concerns the center head. It predicts, for each voxel, an arrow toward the centroid of its lesion. A scalar mask can only say "lesion here, with probability $p$." It cannot say "lesion here, but it is unclear which object this voxel belongs to." In the center field, that second kind of ambiguity appears as locally inconsistent directions. We measure it as one minus the mean resultant length of the arrows in a small window, which is the sliding-window angular variance of GAInS. Our hypothesis is that this clash signal concentrates on three error types: lesions adjacent to vessels, merged small lesions, and transition zones between enhancing rim and necrosis. We further hypothesize that these errors dominate lesion-wise mistakes in post-treatment MRI, so the clash signal should degrade less than softmax entropy or ensemble disagreement under the pre-to-post-treatment shift.
Where This Sits
The directional-field literature supplies the method. Center vector encoding, CDNet, and Cellpose predict arrows toward instance centers and use sharp direction flips to split touching nuclei or cells. Zero Pixel Directional Boundary by Vector Transform regresses arrows toward the nearest boundary and reads off thin edges from their divergence. GAInS computes local angular variance to find overlap regions and reweights the loss there. unMORE uses clashing center directions as an unsupervised signal for splitting objects. All of these treat directional inconsistency as a cue for building a better mask, and all are evaluated on 2D microscopy or natural images.
We reuse their field definitions but change what the inconsistency is for. We treat it as a per-lesion estimate of which-object uncertainty, move it to 3D MRI, and evaluate it as a review-routing score. The vector-transform paper also constrains the design. It assumes boundaries are crisp one-dimensional flips and reports weaker results where boundaries are semantically ambiguous. That assumption is exactly what fails at infiltrative and necrotic margins. This is why we keep the vectors unnormalized: shrinkage then records the ambiguity instead of hiding it.
The calibration work is the most serious competitor. Gradient vector field surgery modifies loss gradients so that Dice-trained models output calibrated probabilities, reducing NLL and ECE "by factors of 4 to 6," including on BraTS-METS. If calibrated probabilities already flag the right lesions, our clash signal is redundant. The experiments are built to answer that question directly.
The mimicry and prospective literatures supply the motivation. BraTS-METS 2023 shows performance collapsing for lesions under 100 mm³, with vascular mimics a known failure mode. Buchner et al. show that necrosis versus recurrence stays hard, and that downstream radiomics falls from a concordance index of 0.77 with expert contours to 0.72 with automated segmentation. Segmentation errors at ambiguous margins therefore propagate into prognosis. The BraTS-METS 2025 Lighthouse benchmark provides 75 cases each annotated four times, which gives us a measured human-disagreement reference. Reader studies such as Meng et al. and Wu et al. show that AI and human errors are complementary. A routing signal is valuable exactly when it points readers at the AI's errors rather than at cases they would have caught anyway.
Research Program
We build one 3D network on the backbone of the BraTS-METS 2023 winning SegResNet. It has three heads: the usual voxel mask, a three-channel boundary-vector head, and a three-channel center-vector head. Both vector heads are trained with unnormalized squared error. This adds six output channels. It does not create the cubic memory blow-up that dense 3D tensor fields would, which the file lists as an open problem.
From the outputs we compute three per-lesion scores:
- the magnitude-band width along the predicted surface, in millimetres;
- the mean clash score over the lesion and a 2 mm shell around it;
- the lesion-level softmax entropy, computed from the same network as a within-model control.
We then run three studies.
The first study tests whether the magnitude width tracks measured human disagreement. On the Lighthouse cases, we check whether the predicted width correlates with the average symmetric surface distance between repeated expert annotations of the same lesion.
The second study tests whether the clash signal flags the right lesions. It is the main result. On the 2023 pre-treatment test data and the 2025 post-treatment data, we label each predicted or missed lesion as correct or erroneous using lesion-wise Dice and detection criteria. Erroneous lesions are sub-typed as edge error, merge, vessel false positive, or rim/necrosis confusion. We then measure how many errors each score captures when a fixed fraction of lesions is flagged for review.
The third study is a frozen-model shift test that stands in for prospective use. We pre-register the model, thresholds, and review budget on pre-treatment data, then apply them unchanged to post-treatment cases. The post-treatment spectrum is enriched for exactly the borderline mimics that curated sets filter out. Degradation from the first to the second setting is the quantity of interest. This is a proxy, not a prospective trial. We state that plainly, and we treat a silent-mode deployment at a partner site as conditional follow-up only if the shift test passes.
A known hazard shapes the analysis. The file notes that internal voids and patchy texture trigger false divergence. Ring-enhancing post-treatment lesions with necrotic cores are exactly this case for the center field. We therefore report the clash score's false-flag rate inside necrotic cores as a separate quantity, and we test a variant that computes clash only in the shell between boundary-field zero crossings.
Baselines
The naive reference is flagging lesions for review at random, at the same budget. A stronger naive reference flags the smallest lesions first. BraTS-METS 2023 shows that lesions under 100 mm³ carry most of the Dice loss, so size alone may capture much of the error. Any learned signal must beat size-ranking to be worth deploying.
The strongest segmentation competitor is the NVAUTO SegResNet that won BraTS-METS 2023. It reported an enhancing-tumor Dice of 0.65 ± 0.25 and lesion detection rates of 76–80% (BraTS-METS 2023). We retrain it on identical splits, with the same augmentation and the same epoch and hyperparameter-search budget as our three-head model. Our segmentation quality must match its lesion-wise Dice within noise. We are not claiming better masks, so a drop in mask quality would count against us.
The strongest routing competitor is the same SegResNet trained with gradient vector field surgery (Lux et al.), with lesion-level entropy as the flag. The file reports calibration gains on BraTS-METS but no Dice or error-capture figure specific to BraTS-METS. We add a five-member deep ensemble and test-time augmentation variance, trained at matched total compute. No reported result for either was found on Lacuna for this task.
Beating these baselines means the clash score captures more lesion-level errors than the best of calibrated entropy, ensemble variance, and size-ranking at 10% and 20% budgets on the post-treatment split, with a paired bootstrap confidence interval that excludes zero.
The key ablations use the same trained network. We compare clash-only against magnitude-only against entropy-only. We compare center-head clash against boundary-head clash. We also compare normalized against unnormalized vector regression, which tests whether conditional-mean shrinkage is what makes the fields informative.
Why not just…?
Why not just use calibrated softmax entropy? For edge-location doubt, this is a genuine equivalent, as the equation above shows. We expect the magnitude signal to tie it. The question is whether "which object" doubt adds anything, and study two settles it with a matched-budget comparison on identical lesions. If calibrated entropy captures vessel and merge errors as well as clash does, the proposal's core claim fails, and we will report it that way.
Why not just use a deep ensemble or test-time augmentation? Ensembles measure disagreement between models. When every member makes the same mistake, as they would on a vessel that consistently looks like a lesion, disagreement is low. The clash signal measures conflict inside a single prediction about object membership, which can be high even when the models agree. Whether that distinction matters in practice is empirical, and the error sub-typing in study two will show which error types each signal catches.
Why not just filter vessel mimics with a Hessian vesselness score? Hessian-based tubular fields work well for vessels, as the Vector Field Transformer shows for retina. This is a real threat for the vessel sub-type specifically. We include a Hessian minor-eigenvector tubularity score as a baseline restricted to vessel false positives. A Hessian score has nothing to say about merges or rim/necrosis confusion, which is where we expect clash to add value.
Why not just build a better segmenter with boundary attention? Boundary-focused designs such as MISNet and DB-MSMUNet improve average Dice, but they still output a mask with no per-lesion doubt. The file also notes that boundary losses overfit a single rater's tracing style exactly where contrast is low. Some errors will remain however good the segmenter is. The deployment question is whether readers can be pointed at those remaining errors.
Decisive Checks
The proposal lives or dies on three results.
The first is on the frozen post-treatment split. Clash must capture significantly more lesion-level errors than calibrated entropy, ensemble variance, and size-ranking at a 10% review budget.
The second is degradation. The drop in error capture from pre- to post-treatment must be smaller for clash than for entropy.
The third is on the Lighthouse cases. The magnitude-band width must correlate positively with measured inter-rater surface distance. Its absence would mean the fields do not encode ambiguity at all.
Two secondary checks follow. Clash false flags inside necrotic cores must stay below the rate of true flags elsewhere. We also recompute the Buchner-style radiomic concordance after correcting only the flagged lesions, to see whether targeted review recovers part of the 0.77-to-0.72 gap between expert and automated contours at a fraction of the reading cost.
References
- Prospective implementation of AI-assisted screen reading to improve early detection of breast cancer
- GAInS: Gradient Anomaly-aware Biomedical Instance Segmentation
- Accurate Nuclear Segmentation with Center Vector Encoding
- CDNet: Centripetal Direction Network for Nuclear Instance Segmentation
- Cell Segmentation in Multi-modality High-Resolution Microscopy Images with Cellpose
- Zero Pixel Directional Boundary by Vector Transform
- unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning
- Beyond scalar losses: calibrating segmentation models via gradient vector field surgery
- The Brain Tumor Segmentation (BraTS-METS) Challenge 2023: Brain Metastasis Segmentation on Pre-treatment MRI
- Radiomics-based prediction of local control in patients with brain metastases following postoperative stereotactic radiotherapy
- Analysis of the MICCAI Brain Tumor Segmentation -- Metastases (BraTS-METS) 2025 Lighthouse Challenge: Brain Metastasis Segmentation on Pre- and Post-treatment MRI
- AI support for accurate and fast radiological diagnosis of COVID-19: an international multicenter, multivendor CT study
- Deep Neural Networks Improve Radiologists’ Performance in Breast Cancer Screening
- Domain Generalization for Retinal Vessel Segmentation with Vector Field Transformer
- Multi-scale information sharing and selection network with boundary attention for polyp segmentation
- DB-MSMUNet: Dual Branch Multi-Scale Mamba UNet for Pancreatic CT Scans Segmentation