Civil engineers learned that how strong a granular material is depends less on how much material it contains than on how its particle sizes are spread. Cement with more total surface area is not stronger. Soil with more fines is not stiffer. A single number summarizing the size spread, called grading entropy, often predicts strength better than any bulk total. Cardiology faces a structurally similar question. Calcium inside an artery-wall plaque is usually summarized as a total volume or a coarse "calcified" label, yet calcified plaques mix a few large plates with many small specks. Plausibly, the specks embedded in a thin fibrous cap matter for rupture more than the total amount. The obvious move is to compute grading entropy of calcified inclusions on intravascular images and correlate it with vulnerability.
That move hides a measurement problem. A size distribution computed from a segmentation depends on how the annotator splits and merges small, faint particles. Two readers who agree on total calcium area can disagree sharply on how many particles there are and how big each one is. Inter-observer variability, the disagreement between experts drawing boundaries on the same image, may therefore hurt distribution-shape descriptors far more than bulk totals. The proposal is to measure that sensitivity first. We then propagate rater disagreement into calcification size-distribution descriptors using a model of individual annotators. Only descriptors that survive this test are asked whether they add vulnerability information beyond total calcium.
Core Claim
Our claim is that particle-size-distribution descriptors of plaque calcification, borrowed from granular mechanics, are only scientifically usable once their between-rater uncertainty is quantified. Quantifying it changes which descriptors are worth studying. Specifically, we expect three things. Descriptors driven by counts and splitting, such as grading entropy and particle count, will show markedly lower inter-rater reliability than total calcified area. A rater-conditioned segmentation model can predict the spread of a descriptor across held-out readers. And a descriptor that clears a reliability threshold will explain expert-labeled fibroatheroma or thin-cap status beyond total calcium, while one that fails will not.
The new element is neither the descriptor nor the rater model alone. It is treating the inter-observer distribution as the object that decides whether a granular-mechanics descriptor carries signal. In the material, granular studies such as Wang et al. and the SEM cement-soil analysis compute particle statistics from segmented microscope images as though the segmentation were exact. Plaque segmentation pipelines, in turn, train only to consensus.
Grounding in Prior Work
The descriptor comes from the particle literature, which also motivates looking past bulk totals. Ke et al. show that cement strength tracks the 32–64 µm fraction rather than total surface area. Xu et al. and Wang et al. find that grading entropy has a strength "sweet spot" and is the most influential microstructural factor in stabilized clay. Zhang Zong-tang et al. link an optimal fractal dimension of grading to minimal deformation under cyclic load. Cyclic load is a fair analogue of the pulsatile stress on a plaque cap.
We reuse their descriptors, not their mechanics. We do not claim plaque calcium behaves like a packed aggregate. Calcified specks sit in a soft matrix rather than forming a load-bearing skeleton, so any link to rupture is an empirical hypothesis. The same literature also constrains us. The SEM cement-soil study notes that going from 2D sections to 3D packing statistics remains empirical. Cross-sectional intravascular frames have the same limitation, so we report per-frame and pullback-aggregated descriptors separately.
The plaque literature supplies the application and the gap. FiAt-Net segments fibrous caps in 3D intravascular optical coherence tomography (IVOCT), a catheter-based imaging method with roughly ten-micron resolution, on the PREDICT cohort. The A-line classification study labels individual radial scan lines as calcified, fibrous, lipid or mixed. The CCTA wall segmentation work bins plaques as soft, mixed or calcified by CT attenuation. All three treat calcium as a homogeneous class and train against a single consensus. None characterizes the size spread of calcified inclusions or how readers disagree about them.
The rater-modeling literature supplies the method. Pionono gives each annotator a learned Gaussian in a small latent space. The mean of that Gaussian captures the annotator's systematic bias and its variance captures their self-inconsistency. A segmentation network conditioned on samples from this space produces coherent alternative masks. The Latent Doctor Model infers an annotator's style from a handful of their past masks and makes the latent prior depend on image difficulty. Disentangling Human Error from Ground Truth instead assumes a single true mask and models raters as pixel-wise confusion.
That last paper conflicts with the other two. It treats disagreement as error to subtract, while they treat it as irreducible. We address the conflict head on by running both kinds of model and asking which one better predicts held-out readers' descriptor values. Faint calcium-to-fibrous transitions in IVOCT may favor the irreducible view, but that is an empirical question. DiFF chooses the "true" mask as the one that maximizes downstream diagnosis. We deliberately avoid that approach. Choosing masks to maximize a vulnerability classifier would make the descriptor-vulnerability association circular.
Research Program
The first component is data. The file lists no public multi-rater plaque dataset, so the program must create a modest one. We take frames from the PREDICT IVOCT pullbacks used by FiAt-Net and from the 33-patient frequency-domain OCT cohort of the A-line study. We sample frames that contain calcium, stratified by whether the frame also contains a fibroatheroma. Four readers delineate every calcified inclusion as a separate object. A subset is re-read after a washout period to measure each reader's consistency with themselves.
Readers mark instances, not just a calcium mask, because splitting and merging are exactly where we expect disagreement. Fibroatheroma and cap labels come from the existing cohort annotations, so the vulnerability endpoint is independent of the new calcium reads.
The second component is the descriptor set. For each frame or pullback segment we compute three bulk quantities: total calcified area, particle count and mean equivalent diameter. We also compute three shape quantities. The first is the area fraction in an intermediate size band, by analogy to Ke et al.. The second is the fractal dimension of the cumulative size curve, following Zhang Zong-tang et al.. The third is grading entropy, written in a simplified form as
$$H = -\sum_{i} p_i \log_2 p_i ,$$
where $p_i$ is the fraction of calcified area falling in the $i$-th size bin and each bin spans a doubling of particle diameter. Inclusions below the imaging resolution are invisible, which truncates the distribution. We therefore compute all descriptors only over resolvable bins and report the lower size cutoff explicitly.
The third component is uncertainty propagation. We train a Pionono-style model that outputs instance masks and is conditioned on rater latents. For each test frame, we sample many plausible annotations from the learned rater distribution and recompute every descriptor on each sample. This yields a predicted distribution of each descriptor rather than a single value. We train a confusion-matrix model in the style of Disentangling Human Error on identical data and compare the two.
The fourth component is the association analysis. We regress vulnerability labels on descriptors at the patient-clustered level. The uncertainty-aware variant carries each descriptor's predicted spread into the regression as measurement error.
Baselines
The naive reference point is total calcified area from a single consensus segmentation, with no shape descriptor and no uncertainty. This is what current pipelines effectively report. Beating it is the whole point. A shape descriptor must add discrimination of fibroatheroma frames on held-out patients, and the added discrimination must exceed the variation across patient-level bootstrap resamples.
The strongest existing segmentation competitor is a deterministic network trained on a single consensus mask, plus a 3D U-Net consensus model for completeness. For IVOCT that is FiAt-Net's backbone family. That paper reports a cap-detection F1 of 81.93% against 69.95% for Swin-Unet on PREDICT. For tissue typing, the A-line study reports 83.47% fibrolipidic-versus-fibrocalcific accuracy. Neither paper reports instance-level calcification segmentation or descriptor reliability: no reported result found on Lacuna for calcified-inclusion size-distribution reliability on IVOCT. The consensus network receives the same frames, augmentation and hyperparameter search budget as the rater model. Every descriptor is computed from its output with the same code.
The strongest uncertainty competitor is a five-member deep ensemble of the consensus network, which performed well in QUBIQ. Top QUBIQ teams averaged Dice between 0.80 and 0.85. On rater modeling itself, Pionono reports a Cohen's kappa of 0.758 on Gleason 2019, where the previous best was 0.750. We give the ensemble the same total training compute as the rater model plus its sampling.
Two ablations isolate the key ingredient. In the first, we keep the rater model but replace instance-level propagation with a single mean mask. This tests whether propagating the full distribution matters. In the second, we keep propagation but remove rater conditioning, using plain test-time dropout. This tests whether rater structure matters. The rater model beats the competitors if three conditions hold. Its predicted descriptor intervals must cover held-out readers' values at close to nominal rates where the competitors under-cover. The rank correlation between predicted spread and actual reader spread must be clearly higher. And it must not lose segmentation Dice relative to the consensus network.
Why not just…?
Why not just report total calcium volume, which is already standard? This is the null hypothesis, and it is a real threat. Bulk calcium may carry all the signal. The granular evidence that totals mislead, from Ke et al. and Zhou et al., comes from load-bearing aggregates, not soft tissue. The association experiment settles this directly. If no reliable shape descriptor improves on total area, we report that negative result.
Why not just have several readers annotate and compute the reliability coefficient directly, with no generative model? For the reliability study itself, that is exactly what we do, and the human ICC is the primary result. The model earns its place only at deployment, where new pullbacks carry one reader or none, as Pionono's discussion of missing rater IDs anticipates. The coverage experiment tests whether the model reproduces the human spread on unseen frames. If it does not, the program reduces honestly to a measurement study.
Why not just use texture radiomics, which already reaches high AUCs for vulnerable plaque? The review in Imaging and Hemodynamic Characteristics of Vulnerable Carotid Plaques cites a radiomic AUC up to 0.988. However, radiomic features are also known to shift with small boundary changes, as shown by Chen et al.. Radiomic features are also not interpretable as size statistics. We include a standard texture feature set as a comparator in both the reliability and the association analyses. If texture features dominate shape descriptors on both, the granular descriptors are unnecessary.
Why not just fuse readers with STAPLE or majority vote and compute descriptors on the fused mask? Fusion averages away exactly the split-versus-merge disagreement we expect to dominate descriptor variance. We test this by including the fused mask as a descriptor source and checking its agreement with each individual reader.
Decisive Checks
The first check comes before any modeling. If grading entropy computed from the human reads has an ICC below 0.5 across readers, we drop it and keep only descriptors that pass. That ICC uses the standard "moderate" cutoff, and its uncertainty is estimated by patient-level bootstrap. The second check is coverage. On frames held out by patient, predicted 80% descriptor intervals should contain held-out readers' values about 80% of the time, and the rater model should clearly beat both the ensemble and dropout. The third check is incremental value. A surviving descriptor must improve held-out discrimination of fibroatheroma frames over total calcified area, with patient-clustered confidence intervals excluding zero. The improvement must also hold after removing frames whose descriptor uncertainty exceeds the between-patient spread.
A transfer check probes the size-distribution premise. Results must replicate on the second OCT cohort, which uses a different system, and must be reported separately for per-frame and pullback-aggregated descriptors. Failure on the first check is itself a useful finding. It would mean that particle-size reasoning imported into plaque imaging measures annotator habits rather than tissue, and the same caution would apply to the SEM-based particle statistics in the granular literature.
References
- Strength Model for Cement-Stabilized Marine Clay: SEM Image Analysis and Microstructural Insights
- Micro-Mechanism of Strength for Cement-Treated Soil Based on the SEM Experiment: Qualitative and Quantitative Analysis
- Preparation of sulfonamide polymer and its effect on cement grinding performance
- Permanent Deformation and Its Unified Model of Coal Gangue Subgrade Filler under Traffic Cyclic Loading
- FiAt-Net: Detecting Fibroatheroma Plaque Cap in 3D Intravascular OCT Images
- Automatic Classification of A-Lines in Intravascular OCT Images Using Deep Learning and Estimation of Attenuation Coefficients
- Automatic Coronary Wall and Atherosclerotic Plaque Segmentation from 3D Coronary CT Angiography
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- The Latent Doctor Model for Modeling Inter-Observer Variability
- Disentangling Human Error from Ground Truth in Segmentation of Medical Images
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge
- Research on the Quality Improvement and Consumption Reduction of Iron Ore Agglomeration Based on Optimization
- Imaging and Hemodynamic Characteristics of Vulnerable Carotid Plaques and Artificial Intelligence Applications in Plaque Classification and Segmentation
- Computed Tomography Radiomic Nomogram for Preoperative Prediction of Extrathyroidal Extension in Papillary Thyroid Carcinoma