Does a Thyroid Grading Model Follow the Grading Rule? Criterion-Level Perturbation Audits

Medical imaging models are moving from yes/no answers ("malignant or not") to finer grades, such as the five thyroid nodule risk levels TR1 to TR5 of the American College of Radiology's Thyroid Imaging Reporting and Data System (TI-RADS). Unlike most diagnostic labels, TI-RADS is defined by an explicit points rule. A radiologist scores five visual criteria (composition, echogenicity, shape, margin, and bright "echogenic foci"), adds the points, and maps the total to a tier. That rule gives something interpretability research rarely has: a ground-truth prediction of how the correct grade should change when one visual property of the image changes. If a nodule gains punctate bright foci (+3 points), the correct grade must rise. If it becomes "taller-than-wide" (+3), the grade must rise. If an unrelated patch of normal thyroid tissue changes, the grade must stay the same.

We propose to use this rule as an answer key for input perturbation, meaning the practice of editing parts of an input and watching the output move. We would make targeted, realistic edits to single criteria in ultrasound images and check whether a model's predicted tier moves in the direction and by the amount the points rule requires. We would pair each edit with a matched "null edit" of equal visual magnitude that should change nothing. The result is a faithfulness score for fine-grained grading that does not depend on whether a heatmap looks plausible. The note tests three hypotheses: that accurate grading models ignore the rule precisely at the intermediate tiers where grading matters most; that aligning images with criterion text improves rule-following; and that rule-following predicts transfer to pathology-confirmed outcomes better than accuracy or saliency scores do.

Core Claim

Current faithfulness checks for fine-grained medical classifiers cannot tell whether a model reads the criteria that define the grade.

Our claim is that criterion-targeted edits with rule-derived expected outcomes, controlled by matched null edits, measure something neither approach can: whether a grading model's decision depends on the clinical criteria in the way the grading system specifies. We expect this measurement to reveal failures that accuracy hides.

Relation to Prior Work

The closest work is Right Diagnoses, Decorative Reasoning: A Perturbation Audit of Medical Chain-of-Thought. It edits clinical premises in multiple-choice questions, such as flipping "denies chest pain" to "reports chest pain", and finds that 72.9% of edits leave both reasoning and answer unchanged. This motivates our work: accuracy and faithfulness come apart in medicine. It also differs in three ways:

  • it works on text multiple-choice questions, not images;
  • it has no ordinal expected outcome, only the question of whether a clinically material edit changes the answer;
  • its authors note that high-acuity shifts were rarely measurable on those benchmarks.

We borrow its operator-battery design and its decoupling-rate idea. We replace "did the answer change" with "did the grade change as the rule says it must."

Two papers supply method:

  • Enhancing Chemical Explainability Through Counterfactual Masking replaces molecular fragments with chemically valid fragments rather than zeros. It shows that zero-masking both leaves the data manifold and leaks information through intact neighbors. We adopt its principle that perturbations should be replacements drawn from real data, not constant fills. Our replacements are ultrasound patches and geometric edits, not molecular fragments.
  • The baseline-fragility results of Evaluating Input Perturbation Methods for Interpreting CNNs and Saliency Map Comparison constrain our design. Any edit can register as a feature in its own right, for example as speckle discontinuities a model might read as foci. This is why every criterion edit has a null twin of equal pixel magnitude.

The audited models come from CMCNet: Aligning Ultrasound Image Embeddings with Textual TI-RADS Representations for Fine-Grained Thyroid Classification. CMCNet aligns ultrasound embeddings to text templates describing the five criteria and predicts the tier from the image alone. Its authors show gains in accuracy but never test whether the encoder uses the criteria.

PCDAL uses flip-consistency as an uncertainty signal for active learning. It measures invariance to edits that should not matter. It cannot measure whether a model responds correctly to edits that should.

Research Program

We work on the Single Thyroid Nodule (STN) dataset used by CMCNet: 600 nodules with transverse and longitudinal views, bounding boxes, and all five criterion labels. Having the criterion labels is essential, because it lets us compute each nodule's point total and hence the tier the rule predicts after any single-criterion edit.

Masks and operators. Operators work inside the nodule. Bounding boxes plus manually traced nodule masks on the evaluation subset define the nodule region. We build three operators we can implement reliably, each with an inverse where one exists:

  • Shape: anisotropic resampling of the nodule region on the transverse view, to make it taller-than-wide (+3 points) or to undo that.
  • Echogenic foci: pasting real punctate-foci patches, harvested from STN nodules labeled with punctate foci, into nodules labeled as having none (+3). The inverse is inpainting foci out with the nodule's own texture.
  • Echogenicity: rescaling nodule intensity relative to the surrounding normal thyroid tissue (parenchyma), crossing the iso/hypo boundary (±1) or the hypo/very-hypo boundary (±1).

Margin and composition edits are harder to make realistically. We treat them as stretch goals rather than promising them.

Null edits. Each operator has a null twin of equal visual magnitude:

  • the same resampling applied to a parenchyma region outside the nodule;
  • a parenchyma patch, not a foci patch, pasted at the same location;
  • the same intensity rescale applied off-nodule.

Expected outcomes. For every edit, the rule gives the expected tier. Some edits cross a tier threshold, for example TR3 (3 points) plus 3 points becomes TR4 or TR5. Others stay within a tier, for example a TR5 nodule at 9 points gaining one more point. These within-tier edits give a second set of "should not change" controls that are clinically meaningful rather than cosmetic.

Edit validation. Two radiologists review a stratified sample of edited images. They rate realism, re-score the edited criterion blind to the intended edit, and try to pick the edited image out of an edited/unedited pair. Edits that fail re-scoring are excluded, so the audit only uses edits that actually changed the intended criterion.

Scores. From these edits we compute three quantities, each reported per source tier:

  • Rule-consistency rate: the fraction of threshold-crossing edits whose predicted tier (or expected tier under the model's softmax) moves in the required direction by at least the required amount.
  • Null sensitivity: the fraction of null and within-tier edits that change the predicted tier.
  • Decoupling rate: the fraction of threshold-crossing edits that leave the prediction unchanged. This is the image-grading analogue of the chain-decoupling rate.

The per-tier breakdown carries the main scientific question: does rule-following hold at TR1/TR2 and TR5 but collapse at TR3/TR4?

Models audited. Plain cross-entropy classifiers on the ultrasound foundation model USFM and on BiomedCLIP backbones; a multitask model with five auxiliary criterion heads in the style of FLaMM-Net; and a reimplemented CMCNet-style center-margin contrastive model aligned to criterion text. Each is trained with several seeds, giving roughly two dozen models spanning a range of accuracies.

Transfer test. We evaluate every model's malignancy score on the pathology-labeled thyroid ultrasound dataset of An ultrasonography of thyroid nodules dataset with pathological diagnosis annotation for deep learning. This tests whether the STN audit predicts transfer to outcomes defined by tissue diagnosis rather than radiologist scoring.

Baselines

Accuracy only. The naive reference is to rank models by in-distribution accuracy, on the implicit assumption that more accurate graders are more faithful. The strongest reported accuracy on STN is from CMCNet: 75.17% balanced accuracy and 91.78% weighted F1 on the five-class task with the USFM backbone. Our reimplementation should land near these figures before we audit it. We will report whether rule-consistency adds predictive information about transfer beyond accuracy, measured as partial correlation with pathology-transfer AUROC after controlling for STN balanced accuracy.

For context on pathology transfer, the dataset paper reports AUROCs of 0.70 to 0.79 for its ThyUS2Path model on external validation. We do not claim to beat that number; we use it to confirm that our models' transfer performance is in a sensible range. We found no reported result on Lacuna for any criterion-level or rule-consistency faithfulness measure on STN, so the audit scores themselves have no published reference point.

Saliency faithfulness, the strongest existing auditing method. The strongest competitor is to score each model with standard saliency measures:

  • the fraction of Grad-CAM and RISE mass inside the nodule box;
  • deletion curves using the most-relevant-first ordering, under both blur and zero baselines, following Brunke et al..

All saliency metrics run on the same models, images and compute, with RISE mask density and count tuned on a held-out split using the same budget we spend on building operators. Our method beats this baseline if rule-consistency predicts pathology transfer more strongly than any saliency metric. It also beats it if rule-consistency identifies tier-specific failures that saliency cannot, for example two models with identical in-nodule heatmap mass but different decoupling at TR3/TR4. We will also report whether the saliency rankings of our models invert between baselines, as Brunke et al. found on ImageNet.

Probing. A linear probe predicts each criterion from frozen embeddings. We give it equal tuning effort and treat its accuracy as the "the model encodes the criteria" alternative.

Null-edit ablation. This ablation isolates our key ingredient. We recompute rule-consistency without null-edit correction and check whether model rankings change. If they do not, the null controls are unnecessary and the audit is simpler than proposed.

Criterion-then-rule reference. A model predicts the five criteria and applies the TI-RADS table. It is rule-consistent by construction, so it gives the ceiling on rule-consistency and shows how much accuracy that ceiling costs.

Why not just…?

Why not just check that Grad-CAM or RISE highlights the nodule? A model can put all its saliency inside the nodule while reading the wrong property of it, such as texture instead of aspect ratio. The neonatal echocardiography staging study, Deep Learning Based Prediction of Pulmonary Hypertension in Newborns Using Echocardiograms, showed plausible 3D Grad-CAM maps. The same model still fell from 0.84 to 0.63 severity F1 on its held-out cohort, so plausible maps did not guarantee robust grading. Brunke et al. also found that deletion-based rankings flip with the baseline. We run saliency as a full baseline, and the transfer-prediction comparison settles which measure is more informative.

Why not just probe the embeddings, or read the multitask heads, to see whether the criteria are encoded? This is a genuine threat. If probe accuracy predicts transfer as well as rule-consistency, our editing machinery is not justified. The conceptual gap is that encoding a criterion is not the same as using it for the final grade. A model can represent "taller-than-wide" linearly and still ignore it when assigning the tier. This is exactly the kind of decoupling Right Diagnoses, Decorative Reasoning documented for text reasoning. The decisive evidence would be multitask models whose shape heads are accurate but whose tier predictions do not move under shape edits. If we do not see such models, we will say so.

Why not just test consistency under flips and noise, as PCDAL does? Flip consistency (PCDAL) measures whether predictions stay put when nothing clinical changes. That is essentially our null-sensitivity score, and we report it. It cannot reveal a model that is stable everywhere because it ignores the criteria: such a model scores perfectly on flips and fails our threshold-crossing edits.

Why not just deploy the criterion-then-rule model and skip auditing? If this model matched end-to-end accuracy, that would be the right clinical answer, and our experiments will report it. CMCNet's comparison against FLaMM-Net-style multitask heads suggests end-to-end aligned models are more accurate. In addition, radiologist criterion labels are subjective, so a hard-coded rule inherits their noise. Audits are needed for the end-to-end models people actually train.

Decisive Checks

Edit validity. The first check is whether the edits are valid. If radiologists can reliably tell edited from real images, or fail to re-score the intended criterion, the operators are measuring artifact response and must be fixed before anything else is interpreted. The null-edit sensitivity numbers give a second, automatic version of this check.

Tier-specific decoupling (H1). This hypothesis is confirmed if decoupling rates at TR3/TR4 exceed those at the extreme tiers by a margin larger than the seed-to-seed spread, in models whose per-tier accuracy is comparable. If decoupling is uniform across tiers, the "granularity gap" story is wrong and we will report the audit as a general faithfulness tool instead.

Text anchoring (H2). This compares CMCNet-style models with cross-entropy and multitask models at matched balanced accuracy. It asks whether text anchoring buys rule-following, or only accuracy.

Transfer prediction (H3). This is the main test of usefulness. Across roughly two dozen models, we correlate each candidate metric with pathology-transfer AUROC on the Hou et al. dataset: STN accuracy, probe accuracy, saliency scores under each baseline, and rule-consistency with and without null correction. If rule-consistency does not beat saliency and probing here, the proposal's practical claim fails, even if the per-tier findings stand as a descriptive result.