Measure the Bite Before Writing the Report

TL;DR for operators

A clinical report-generation system does not necessarily need a large language model to interpret every input and compose every sentence. When relevant physical relationships can be measured directly, deterministic computation can provide the evidence, conventional machine learning can predict selected clinical categories, and standardized templates can produce the report.

In Occlusal Geometry in Closed Form for Orthodontic Report Generation, George and colleagues demonstrate this approach using registered three-dimensional dental scans.¹ Their system extracts 31 geometric features, selectively predicts 13 report fields, and generates a structured narrative. It improves held-out text-similarity scores over a corpus-prior baseline and places third in the ODIN 2026 Bite2Text test phase.

The operational lesson is to assign each reporting decision to the method supported by its evidence, not automatically to the most flexible model. Some fields remain poorly predicted, and a stronger text-similarity score does not guarantee greater clinical correctness. The appropriate deployment target is clinician-reviewed documentation, with field-level validation and explicit handling of uncertain findings.

A clinical report begins with measurements, not sentences

Consider an orthodontist examining upper and lower three-dimensional dental scans. The report may describe overbite, overjet, crossbite, arch constriction, crowding, and sagittal relationships between teeth.

These observations are not equally difficult to obtain. Some describe distances or spatial overlaps visible in the scans. Others depend on individual tooth relationships, photographic evidence, or clinical judgment.

Treating the entire task as image-to-text generation gives one learned model responsibility for measurement, interpretation, and wording. Errors in those operations can become difficult to distinguish once they appear in a fluent clinical paragraph.

The Bite2Text dataset offers a different starting point: upper and lower scans are already registered in occlusion, meaning their relative positions are available. The system can therefore measure relevant relationships before deciding how to describe them.

The released corpus contains 996 cases across overlapping report families. The authors evaluate structured predictions and generated narratives against clinician-authored reports, while examining how variations in clinical documentation affect automated scoring.

Reconstructing the anatomical frame makes measurement possible

A geometric measurement is only meaningful when the coordinate system is correct.

Although the scan documentation describes a particular anatomical orientation, the released data do not consistently follow it. Assuming that the mesh axes already represent the intended directions can silently corrupt distances and relationships.

The proposed pipeline first reconstructs each patient’s anatomical frame. It uses upper and lower arch centroids to identify the superior direction, arch taper and mandibular closure to infer the anterior direction, and tooth-number annotations from the auxiliary Bits2Bites dataset to resolve left and right.

The authors find that restricting the closure cue to the mandible improves orientation detection and downstream field accuracy. This is a preprocessing correction, but it directly affects the reliability of every subsequent measurement.

Once oriented, the arches are represented through angular ridge profiles around a shared mandibular reference point. A one-degree grid spanning -150 to +150 degrees captures occlusal extremes and buccal and lingual boundaries.

From these profiles, the system computes 31 features covering vertical and horizontal bite relationships, transverse overlap, crossbite extent, cusp interdigitation, arch width, occlusal curves, and anterior irregularity. Geometrically implausible measurements are marked missing instead of being treated as valid observations.

An auxiliary validation on 200 annotated patients provides evidence that the measurements capture meaningful clinical differences. Geometry-derived features support anterior-bite classification accuracy of 0.835 against a 0.405 majority baseline, and transverse-pattern accuracy of 0.820 against 0.700.

These are categorical validation results, not proof of millimeter-level clinical measurement accuracy. The main Bite2Text corpus does not provide direct numerical ground truth for the geometric measurements.

Each predicted field must outperform its own baseline

Having 31 measurements does not make every report field reliably predictable.

The authors train separate histogram-based gradient boosting classifiers for 13 structured clinical fields. Each classifier competes against a simple alternative: always predicting that field’s most frequent class in the training corpus.

A learned predictor is deployed only when patient-level cross-validation demonstrates higher accuracy than its field-specific majority baseline. Otherwise, the system uses the modal value.

The results show why this restriction matters. Constriction, overjet, crossbite, and overbite improve by approximately 0.16 to 0.25 accuracy points over their respective baselines. Sagittal classifications gain approximately 0.09 to 0.13, while midline and upper-arch crowding fail to demonstrate improvement.

This is selective model deployment, not selective certainty. A modal prediction can improve aggregate accuracy when a learned classifier performs poorly, but it supplies no patient-specific evidence for that field.

The held-out scan evaluation shows the value and remaining limitations of the resulting architecture.

Source: Section 4, Table 1; 296 patient-disjoint intraoral scan cases. Oracle fields use ground-truth structured labels rather than predicted fields.

Geometry produces an absolute improvement of 0.0276 in BLEU-4 and 0.0654 in METEOR over the prior-only renderer. The oracle configuration demonstrates how much additional captioning performance is possible with more accurate structured fields under the same evaluation setup.

The deterministic renderer then converts field values into a standardized six-part clinical narrative. It separates the problem of estimating findings from the problem of expressing them, making it easier to inspect which component supplied a questionable statement.

Photographs add information only when they contribute distinct evidence

Arch-level geometry is less effective for findings requiring detailed tooth relationships or visual assessment. The reported accuracy of weaker sagittal fields remains around 0.53 to 0.59.

The authors therefore investigate a complementary photographic model using a shared ConvNeXt-Tiny backbone and five standardized intraoral views. Field-specific fusion weights determine how geometric and photographic predictions are combined.

On a disjoint held-out evaluation, fusion improves mean structured-field accuracy from 0.601 to 0.683. It also raises end-to-end BLEU-4 from 0.2810 to 0.2930 and METEOR from 0.5079 to 0.5213.

These are post-challenge results. The photograph-fusion improvement was not included in the officially submitted third-place system.

An input-quality test gives a second lesson. Repeating one available photograph to fill missing view slots reduces photograph-model accuracy from 0.641 to 0.508. The pipeline consequently disables photographic prediction when fewer than three distinct views are available.

Additional model inputs are beneficial only when they contain information that the model can use reliably. Duplicating a view changes input format without adding observational evidence.

Text similarity can reward a less informative clinical report

The paper’s evaluation analysis challenges a common interpretation of generated-report benchmarks: that stronger agreement with reference wording implies stronger clinical correctness.

Across 498 paired clinician reports, token Jaccard similarity is 0.472 and structured-finding Jaccard similarity is 0.469. Clinicians frequently describe the same case differently.

One comparison is especially revealing. A constant corpus-medoid report, containing no patient-specific geometric reasoning, achieves a reconstructed captioning mean of 0.547 against real scan references. A genuine second-clinician report achieves 0.382 against the first clinician’s report in the reported comparison.

The constant report has not demonstrated superior clinical quality. Its advantage reflects how common wording and reference conventions interact with automated evaluation.

The authors also reconstruct the deployed METEOR-lite scorer. Its F-mean weights recall much more strongly than precision:

Here, $P$ is matched-token precision and $R$ is recall. The implementation additionally penalizes fragmentation of matched sequences. Its recall-heavy weighting makes missing expected language more expensive than including some extra unmatched content.

The resulting incentive can favor standardized, comprehensive-sounding reports even when additional assertions are not supported by patient-specific evidence.

A locally reproduced factuality evaluation exposes another divergence. Under a substitute Qwen2.5-7B-Instruct RadFact judge, the prior-only configuration reaches F1 of 0.468, slightly above geometry’s 0.459. Geometry nevertheless achieves a higher captioning score and slightly higher locally calculated combined score.

This comparison is diagnostic of evaluation disagreement, not proof that the prior-only system is clinically preferable. The substitute judge also prevents direct numerical comparison with official RadFact results.

Competitive CPU deployment does not establish diagnostic autonomy

The submitted system placed third in the ODIN 2026 Bite2Text test phase, with BLEU-4 of 0.2680 and METEOR of 0.4629, within 0.022 BLEU-4 of the first-ranked entry.

These hidden-test scores are separate from the held-out results discussed earlier.

The gradient-boosting model bundle occupies approximately 5 MB, and tested CPU inference completes in 3.1 to 9.3 seconds per case. The implementation also handles incomplete and alternative input layouts and addresses excessive OpenMP overhead during single-case prediction.

For clinical documentation product teams, Cognaptus infers a useful design pattern: measure directly observable properties, validate learned predictors independently for each field, gate complementary modalities according to input quality, and render reports through auditable templates.

This approach can limit dependence on expensive generative inference for suitable structured tasks. However, the paper does not establish comparative production costs, prospective clinical effectiveness, or safety in routine orthodontic practice.

The boundaries are substantive. Individual cusp-to-groove relationships remain difficult to recover from arch-level profiles; some predictions revert to corpus defaults; photo-derived health labels partly reflect clinicians’ reporting choices; and numerical measurement ground truth is missing.

A production workflow should therefore expose field provenance, distinguish measurement from prediction and default values, test clinical factuality independently of text similarity, and require clinician review before findings become part of the patient record.

The system’s central contribution is choosing who makes each decision

The study shows that competitive clinical report generation can emerge from a division of computational responsibilities.

Geometry handles measurable spatial relationships. Validated statistical models handle selected uncertain classifications. Photographs contribute when their distinct evidence is available and useful. Deterministic rendering produces consistent documentation.

The reported gains establish the feasibility of that architecture within a registered-scan benchmark. They do not eliminate uncertainty or turn fluent reports into verified diagnoses.

For developers, the more general design principle is to allocate decision authority according to observability, validation evidence, and error consequences. A model should receive responsibility for a clinical field because it has demonstrated that it can improve the decision, not simply because the model can generate an answer.

Cognaptus: Automate the Present, Incubate the Future.