Radiologists already mark most tumors in routine care. They draw two perpendicular lines across each lesion, a RECIST measurement, rather than tracing its outline. Several methods now train lesion segmenters from these lines alone. They use losses that concentrate training on the uncertain ring between a shape that is surely inside the lesion and a shape that surely contains it. They reach Dice overlap within a point of full supervision. What nobody has checked is whether the probabilities these models output at the lesion margin can be trusted. That matters because clinical decisions rest on quantities computed from segmentations, such as tumor area or the ratio of cyst to solid tumor, and on how confident we can be in those numbers.
Our proposal changes what the ring loss asks the model to learn. Instead of forcing hard agreement in the ring, the loss pins the expected lesion fraction in each distance band of the ring to a frequency estimated from a few dozen densely traced lesions. The model stays free to decide which margin pixels belong to the lesion, but its probabilities are anchored to how often margin pixels actually are lesion. We then judge the result by a clinical yardstick rather than Dice alone: how wide a statistically valid confidence interval on a downstream biomarker must be, both in-distribution and after moving to another hospital's data.
Core Claim
Weakly supervised lesion segmenters with region-restricted losses are likely miscalibrated exactly at the lesion margin, where volumetric biomarkers are decided. CARE builds confidence intervals for ratio biomarkers without retraining, and it shows that segmentation miscalibration, not estimation noise, dominates biomarker uncertainty. Interval methods like CARE keep their nominal coverage on exchangeable data, so miscalibration does not show up as lost coverage. It shows up as wider intervals. We claim that a margin-calibrated weak loss produces biomarker intervals that are clearly narrower than those of current weak methods at equal coverage. We further claim it closes most of the width gap to full supervision, while Dice stays at the level of label-space co-training.
The three input directions play different roles. The region-weighted loss is the mechanism. The annotation bottleneck is the constraint: only RECIST lines plus a very small dense calibration budget. The clinical validation gap supplies the evaluation, through biomarker interval width, coverage under a hospital shift, and sensitivity near clinical thresholds.
What exists and what is missing
Two papers already combine RECIST supervision with region-restricted losses. Tang et al.'s Regional Level Set loss confines a boundary-energy term to a box around the RECIST marks and reports 92.6% Dice on DeepLesion. Zhou et al. build an inner quadrilateral $Q$ and a circumscribed circle $C$ from the line endpoints and co-train two networks to agree inside the ring $C \setminus Q$, reaching 0.907 Dice on KiTS19. Both push predictions toward crisp binary decisions in exactly the region where the truth is least known, and neither reports calibration or any downstream quantity. These papers supply our geometry and our strongest competitor. Our proposal keeps their ring and replaces the agreement target.
Lee et al. show in a generative setting why region weighting is needed at all. Uniform losses dilute gradients from small lesions, and their lesion-weighted post-training cut lesion error by 12.4%, but only by using dense expert masks. They motivate the mechanism. The pipelines of Agarwal et al. and Can et al. rely on GrabCut, random walks and conditional random fields to turn weak marks into masks. Can et al. found that uncertainty filtering, not the random field, was what stabilized training. That constrains us: the margin uncertainty should be modeled explicitly rather than smoothed away.
On the validation side, de la Rosa et al. set the bar for external testing, with 1,686 scans from 11 scanners, biomarker correlations and a reader study. They did this with fully supervised ensembles. The label-free metric estimation work shows that confidence-based monitoring degrades under shift, and CARE states that its guarantees assume matched calibration and test distributions. Both motivate testing our intervals under a hospital shift rather than only in-distribution.
Research Program
For each RECIST-annotated lesion we build $Q$ and $C$ as in Zhou et al. For each pixel in the ring we assign a normalized position $t \in [0,1]$: 0 on the boundary of $Q$, 1 on the boundary of $C$, measured along the ray from the lesion center. We split the ring into $B$ bands by $t$ (we expect $B \approx 8$). From $K$ densely traced lesions we estimate $q_b$, the fraction of pixels in band $b$ that truly belong to the lesion. We plan $K$ between 10 and 50, drawn from DeepLesion's 1,000 traced lesions or from KiTS training masks. The $q_b$ may also be conditioned on coarse lesion size, since small lesions have proportionally thicker margins.
The training loss has three parts. Pixels inside $Q$ and outside $C$ get ordinary cross-entropy with full weight. Ring pixels get the image-driven consistency term from Zhou et al. at reduced weight. A band-wise expected-area term ties the model's average ring probability to the calibration frequencies:
$$\mathcal{L}_{\text{band}} = \sum_{b=1}^{B} \Big( \tfrac{1}{|R_b|}\sum_{i \in R_b} p_i - q_b \Big)^2 ,$$
Here $R_b$ is the set of ring pixels of one lesion in band $b$, and $p_i$ is the predicted lesion probability at pixel $i$. The term constrains aggregate frequency, not individual pixels. So the network can still follow image edges to choose which margin pixels are lesion. However, it cannot drive every ring probability to 0 or 1 unless the evidence supports it. Because expected area is exactly what volumetric and ratio biomarkers sum over, this term targets the error that matters clinically.
We also add a spatially smooth weight that decays with $t$ instead of a step. This addresses the boundary-artifact problem that step-function weights cause in region-weighted losses.
Downstream, we feed each trained model's probability maps into CARE unchanged. We then measure interval width and coverage for two biomarkers: lesion area on the key slice, and the cyst-to-mass ratio on KiTS23 key slices. This isolates how the training loss affects biomarker reliability.
Baselines
The naive reference is a standard nnU-Net-style segmenter trained with uniform Dice plus cross-entropy on masks produced offline by GrabCut from the RECIST lines. Zhou et al. report 0.891 Dice for GrabCut pseudo-labels on KiTS19. Tang et al. report that nnU-Net trails their method by 2.6 Dice points on DeepLesion, at about 90.0% against 92.6%. It is run on the same backbone and the same number of epochs as our method.
The strongest existing method on this question is label-space co-training with the ring consistency loss, which reports 0.907 Dice on KiTS19 against a fully supervised 0.912. We reimplement it with the same Swin backbone, data splits and hyperparameter search budget as our method, since no public calibration or biomarker numbers exist for it. The Regional Level Set loss (92.6% Dice on DeepLesion) is a second strong competitor on DeepLesion, run the same way.
The fully supervised model trained on dense masks gives the upper bound on interval width. No reported result was found on Lacuna for CARE interval widths on KiTS23 or BraTS; the paper reports only that its intervals are tighter than those of the ACQR baseline. We therefore recompute all interval widths ourselves.
The key ablation removes $\mathcal{L}_{\text{band}}$ but keeps everything else, including the smooth weights. A second ablation replaces the fitted $q_b$ with a fixed linear ramp that uses no dense masks. Together these separate the effect of the calibration data from the effect of the soft-target geometry.
Our method beats the strongest baseline if two conditions hold. First, it matches co-training's Dice within about half a point. Second, it reduces CARE interval width on the area biomarker by a margin larger than seed-to-seed variation, at equal nominal coverage, and also beats every post-hoc recalibration baseline described below.
Why not just…?
Why not just train co-training as published and apply temperature scaling on the same $K$ dense lesions? This is the most serious threat, and we treat it as a primary baseline, not a straw man. A single global temperature cannot fix miscalibration that is concentrated in the margin while the lesion core and background are already confident, but a per-band or distance-conditioned temperature might. We therefore also fit a band-conditioned post-hoc recalibration on the identical $K$ lesions. If it matches our interval widths, the training-time term adds nothing. In that case we would report the result as a finding that weak region losses need only post-hoc margin recalibration. The experiment decides this directly, because both approaches use exactly the same dense-label budget.
Why not just run CARE on any weak model, since it keeps coverage anyway? Coverage is preserved in-distribution, and that is exactly why width is the right metric. CARE's authors show that miscalibration inflates biomarker uncertainty, so a poorly calibrated weak model pays in wide intervals. That shows up as more cases flagged as ambiguous near decision thresholds, which is a real clinical cost. Under the hospital shift, CARE explicitly loses its guarantee. Measuring which training loss degrades coverage least after the shift is therefore part of the question, not something CARE already answers.
Why not just spend the dense-label budget on semi-supervised training? With $K$ dense lesions, one could fine-tune the co-trained model on them directly. We include this baseline at identical $K$. Our expectation is that 10 to 50 masks improve Dice marginally but barely move margin calibration across thousands of weakly labeled lesions. The band term uses those same masks to constrain all of them. If direct fine-tuning matches us, the band term is unnecessary.
Why not just use an ensemble? De la Rosa et al. show that ensembles generalize well across scanners, but they trained on dense masks and ran three 3D networks. We include a three-seed ensemble of the co-training model as a calibration baseline, since ensembling often improves calibration. We report its compute cost next to its interval widths.
Proposed Experiments
We run the main in-distribution study on KiTS19/KiTS23 with simulated RECIST lines on key slices, following Zhou et al. Dense masks provide ground-truth biomarkers there, and KiTS23's cyst labels give a ratio biomarker matching CARE's setting. A second in-distribution study uses DeepLesion. Models train on RECIST bookmarks, the $K$ calibration lesions come from part of the 1,000 traced lesions, and the rest form the test set. Because DeepLesion ignores unannotated co-existing lesions, we score only bookmarked lesions to avoid the missing-annotation confound.
For the shift test, we train on DeepLesion and evaluate on KiTS kidney tumors converted to RECIST lines, a different institution and lesion population. CARE is calibrated on DeepLesion and tested on KiTS. We report empirical coverage and interval width for every method.
The metrics are Dice, margin-restricted expected calibration error (computed within 3 mm of the true boundary), CARE interval width and empirical coverage for both biomarkers, and sensitivity in flagging lesions near an illustrative area threshold.
Three results decide the question. The first is whether co-training and the level set loss show markedly worse margin calibration than full supervision at similar Dice. If not, the premise is wrong and we stop. The second is whether our loss narrows interval width beyond band-conditioned post-hoc recalibration at the same $K$. The third is whether coverage under the DeepLesion-to-KiTS shift degrades less for our model than for the baselines. We also sweep $K$ over 0 (the fixed ramp), 10, 25 and 50 to show how much dense annotation the calibration actually requires. That curve is the practical answer to the annotation bottleneck.
References
- CARE: Confidence-aware Ratio Estimation for Medical Biomarkers
- RECIST Weakly Supervised Lesion Segmentation via Label-Space Co-Training
- Weakly-Supervised Universal Lesion Segmentation with Regional Level Set Loss
- Lesion-Aware Post-training of Latent Diffusion Models for Synthesizing Diffusion MRI from CT Perfusion
- Weakly Supervised Lesion Co-Segmentation on Ct Scans
- Learning to Segment Medical Images with Scribble-Supervision Alone
- A Robust Ensemble Algorithm for Ischemic Stroke Lesion Segmentation: Generalizability and Clinical Utility Beyond the ISLES Challenge
- Label-free estimation of clinically relevant performance metrics under distribution shifts