When heart muscle loses its blood supply, the oxygen-starved tissue dies and is replaced by stiff collagen scar. Doctors see this scar on late gadolinium enhancement (LGE) MRI, a scan taken minutes after a contrast agent is injected, where scar shows up bright. Clinically, the scar is usually summarized by its volume. Work in oncology suggests that a lesion's three-dimensional shape (how flat it is, how ragged its surface is, what happens at its rim) can carry more prognostic signal than its size. Shape, though, is exactly what human readers disagree about most. Two cardiologists using two accepted intensity-threshold rules will draw different scar borders, and the bright-but-not-too-bright rim is where they diverge. Any shape number computed from one segmentation therefore quietly inherits one reader's convention.
This note proposes treating each geometric scar descriptor as a distribution over plausible reader conventions rather than as a single number. We train a 3D segmentation model that can be asked "draw this scar the way a 5-standard-deviation reader would" or "the way a half-maximum reader would." From the resulting family of masks we measure, for every shape descriptor, how much of its variation across patients is real and how much is convention. We then test a specific biological hypothesis: the width of the rim where conventions disagree is not noise, and it tracks how severe the oxygen-deprivation injury was. In this note that severity is indexed by microvascular no-reflow, meaning regions where blood fails to return even after the artery is reopened.
Core Claim
There are two linked claims. The first is a measurement claim. Three-dimensional geometric descriptors of ischemic scar differ sharply in how robust they are to reader convention. A segmentation model conditioned on an explicit convention variable predicts the spread of each descriptor across real human readers better than generic uncertainty methods do (dropout sampling, test-time augmentation, random contour jitter), because reader disagreement in LGE is structured: it comes mainly from the choice of threshold rule and of reference region. It is not random boundary noise.
The second is a biological claim. The convention-disagreement band, the shell of voxels that some conventions call scar and others do not, is an informative feature in its own right. Its thickness predicts no-reflow beyond what scar volume predicts. If the first claim holds and the second fails, the work still delivers a reliability map for scar shape features. If both hold, rater disagreement becomes a measurement rather than something to average away.
Where This Sits
The geometric side motivates the proposal. Hao et al. isolated the outer voxel "shell" of tumors on PET and predicted distant failure better than interior texture and volume features. Wu et al. found flatness to be the main shape predictor of renal malignancy. A pairwise ML recurrence platform found least axis length beat volume and intensity for esophageal recurrence. None of these studies concerns fibrosis, and all compute features on one manual mask. Chen et al. built a thyroid CT nomogram from shape and texture features and named manual segmentation variability as the main barrier to reproducibility. That is the constraint we take seriously.
On the scar-segmentation side, three systems supply our backbones and comparison points. ScarNet pairs a MedSAM transformer encoder with a U-Net decoder and reaches median scar Dice 0.912 on the DETERMINE registry. Its labels are a fixed two-cardiologist consensus drawn with the half-maximum rule. Zhou et al. enforce the containment no-reflow ⊂ infarct ⊂ myocardium on EMIDEC, against a single consensus. MTI-MyoScarSeg segments scar from motion in non-contrast cine MRI, but it transfers LGE labels without quantifying their uncertainty. Our proposal does not try to beat these systems on Dice. It asks what their masks mean geometrically once the reference label is no longer treated as unique. The fibrosis segmentation review gives the mechanism behind our design: it attributes manual LGE disagreement largely to the arbitrary choice between n-SD and half-maximum thresholding and to partial-volume effects at thin walls. Disagreement of that kind is structured, which is why a convention variable should capture it.
The rater-modeling methods supply the technique. Pionono gives each rater a Gaussian in an 8-dimensional latent space, with the mean as systematic bias and the covariance as the rater's inconsistency with themselves. Liu et al. use an unconditioned 6-dimensional latent variable to sample diverse masks. Both work in 2D and on non-fibrotic targets, and neither propagates samples into shape features or outcomes. DiFF is the closest conceptual neighbor. It also argues that majority voting destroys diagnostic boundary information. However, it resolves disagreement by weighting raters toward whatever helps a particular classifier. We keep the disagreement and measure it, so our results do not depend on any one downstream network.
Research Program
The first component is a convention-labeled training set. Real multi-reader cardiac scar data are scarce, and the file lists no public multi-rater LGE scar dataset. We therefore build "synthetic readers" that reproduce the documented sources of disagreement. Within the expert myocardium mask of each EMIDEC case, we generate scar masks under several threshold rules: 2, 3 and 5 standard deviations above remote myocardium, and half-maximum. For each rule we also perturb the remote-myocardium reference region. This yields a family of labels per case, each tagged with its convention. Separately, we propose collecting a small real multi-reader test set: three to four readers independently re-contour a held-out subset of roughly 30 EMIDEC cases. This set is never used for training. Its only purpose is to check whether synthetic conventions predict real human spread.
The second component is the reader-conditioned 3D segmenter. We adapt the Pionono latent design to a 3D nnU-Net-style backbone and keep the containment penalty of Zhou et al. so that sampled masks remain anatomically valid. Each convention gets a learned latent distribution. Real readers are later placed in the same latent space with a few fitted cases each. Sampling from the latent space yields a set of plausible 3D masks per patient.
The third component is geometry under uncertainty. For each sampled mask we compute volume, surface-to-volume ratio, flatness, least axis length, transmurality (the fraction of wall thickness the scar crosses), and a Hao-style shell intensity feature. We also compute one new quantity, the disagreement band. Following Jin et al., who turned masks into signed distance maps, we convert each sampled mask into a signed distance field. The band thickness is then the per-patient spread of the zero-level surface across samples, summarized as mean and 90th-percentile distance in millimetres. For every descriptor, reliability is the intraclass correlation (ICC), the share of total variance that lies between patients rather than between conventions within a patient.
Baselines
The naive reference point is a single deterministic 3D nnU-Net trained on the EMIDEC consensus, with shape features computed on its one output mask. Its convention spread is zero by construction. This is what current radiomics pipelines do, and it tells us how much a single-mask analysis overstates its own reliability.
The strongest segmentation competitor is the containment-constrained U-Net of Zhou et al., which reports infarct Dice 0.747 and no-reflow Dice 0.538 on EMIDEC. We reimplement it on the same folds and do not expect to beat it. The requirement is that our convention-conditioned model, evaluated at the latent setting nearest the consensus, stays within a few Dice points of it. Otherwise our shape distributions come from a worse segmenter. ScarNet's 0.912 median scar Dice was reported on DETERMINE, not EMIDEC. If DETERMINE access is obtained we will use ScarNet as the backbone there. Otherwise we cite the number only for context, since it is not comparable across datasets.
The strongest competitors for predicting feature spread are cheaper uncertainty methods: Monte Carlo dropout, test-time augmentation, contour jittering in the spirit of Jin et al., and the unconditioned variational model of Liu et al.. All of them run on the same backbone, data and folds, with matched sample counts and the same hyperparameter search budget. The metric is how well each method's predicted per-patient descriptor spread matches the spread observed among real readers on the held-out set, measured by rank correlation and by interval coverage. We have no reported result on Lacuna for any method predicting reader spread of scar shape descriptors. Beating these baselines means higher spread correlation and coverage closer to nominal, with confidence intervals bootstrapped over patients that exclude a tie.
The key ablation removes the convention label: the same latent model trained on the pooled synthetic masks without tags. If it matches the conditioned model, explicit convention structure is not what carries the result.
For the biology claim, the reference is a logistic model predicting no-reflow presence, and a regression predicting no-reflow extent, from infarct volume alone. We compare it against volume plus point-estimate shape features, and against volume plus distributional features, which include band thickness. All features are computed from the outer infarct surface only, so no-reflow voxels never enter. No reported result was found on Lacuna for predicting no-reflow from infarct geometry.
Why not just…?
Why not just compute shape features on one good automated mask, since deep networks are already deterministic and near expert level? Being deterministic makes a measurement repeatable, but it does not make it valid. A network trained on half-maximum labels reproduces half-maximum geometry every time. Chen et al. and the fibrosis review both point to convention dependence as the reproducibility problem. The naive baseline makes the cost concrete: we report how many descriptors fall below an acceptable ICC across real readers even though the single mask shows zero variance.
Why not just use dropout or test-time augmentation for uncertainty? This is the real threat. If generic sampling predicts reader spread as well as convention conditioning, the extra machinery is unjustified. Our case rests on the review's account that LGE disagreement is systematic, a choice of rule rather than random noise. Dropout mostly reflects the model's own uncertainty, not systematic reader bias. The spread-prediction comparison settles this directly. If the cheaper methods tie, we will recommend them and keep only the reliability analysis.
Why not just collect three to five readers and use majority vote or STAPLE? Consensus labeling gives a cleaner single mask but by design deletes the band we hypothesize is informative. DiFF found that majority voting washes out diagnostic boundary structure. We include a consensus-mask arm in the no-reflow experiment, so if the band adds nothing, consensus wins and we say so.
Why not just use scar volume? Volume may well capture most of the signal. The shell and least-axis findings (Hao et al., the esophageal platform) come from tumors, not scar. The nested-model comparison tests whether geometry has incremental value, with likelihood-ratio tests and cross-validated AUC.
Proposed Experiments and Decisive Checks
The first experiment checks that segmentation quality is preserved. We use five-fold cross-validation on EMIDEC's 150 exams and compare against the reimplemented Zhou model at the consensus-nearest latent setting. The second experiment measures reliability and spread prediction on the held-out real-reader subset. It reports per-descriptor ICC under each method and the correlation between predicted and observed spread. This is the decisive test of the measurement claim, and it fails if the unconditioned ablation or dropout matches the conditioned model.
The third experiment tests the biology claim on EMIDEC through the nested no-reflow models. It fails if band thickness adds nothing beyond volume. As a guard against a trivial explanation, we also regress band thickness on infarct volume and contrast-to-noise ratio, to confirm that the band is not just a proxy for size or image quality.
The fourth experiment is exploratory because its cohort is small. On the 50-patient cine-plus-LGE cohort of MTI-MyoScarSeg, we test whether segments inside the disagreement band show intermediate motion deficits in the Motion-Net displacement fields, between core scar and remote tissue. Such a pattern would support the idea that the band marks partially stiffened, transitional tissue. If DETERMINE registry follow-up endpoints are accessible (the file does not describe them), the same distributional features would then be tested against outcomes. That step is contingent and is not part of the core claim.
References
- Shell feature: a new radiomics descriptor for predicting distant failure after radiotherapy in non-small cell lung cancer and cervix cancer
- Interpretable multiphasic CT-based radiomic analysis for preoperatively differentiating benign and malignant solid renal tumors: a multicenter study
- Pairwise machine learning-based automatic diagnostic platform utilizing CT images and clinical information for predicting radiotherapy locoregional recurrence in elderly esophageal cancer patients
- Computed Tomography Radiomic Nomogram for Preoperative Prediction of Extrathyroidal Extension in Papillary Thyroid Carcinoma
- ScarNet: a novel foundation model for automated myocardial scar quantification from late gadolinium-enhancement images
- Anatomy Prior Based U-net for Pathology Segmentation with Attention
- Contrast-Free Myocardial Scar Segmentation in Cine MRI using Motion and Texture Fusion
- Recent Advances in Fibrosis and Scar Segmentation From Cardiac MRI: A State-of-the-Art Review and Future Perspectives
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- Variational inference for quantifying inter-observer variability in segmentation of anatomical structures
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Deep Esophageal Clinical Target Volume Delineation Using Encoded 3D Spatial Context of Tumors, Lymph Nodes, and Organs At Risk