PaperlayerView exact revision on arXiv

Source-linked AI summary

EviDent-CBCT: Evidence-Bottlenecked Report Generation from Dental CBCT under Non-Exhaustive Report Supervision

Ruiyang Hao, Zhi Qin Tan, Yulan He, Owen Addison, Yunpeng Li

arXiv:2610.02375v1cs.CVcs.AI

TL;DR

Routine CBCT reports may omit image findings, making it difficult to learn report generation when non-mention does not establish absence. EviDent-CBCT predicts a discrete evidence record and uses it as the sole case-specific input to an image-blind report generator. Across three validation runs, it outperformed the strongest controlled direct generator on RadFact-Lite-Dental logical-F1 and placed second in automated and third in blinded clinical ODIN 2026 evaluation.

  • Problem

    Routine clinical CBCT reports do not exhaustively document image findings, and a non-mention may indicate absence or non-reporting.

    Source: §1 · p. 1

  • Method

    EviDent-CBCT predicts tooth-level, global, and tooth-IAC evidence, applies dental-logic consistency projection, and supplies the resulting discrete record to an image-blind report generator.

    Source: §1 · p. 1

  • Results

    Across three validation runs, EviDent-CBCT achieved 0.402 ± 0.003 RadFact-Lite-Dental logical-F1 versus 0.371 ± 0.018 for the strongest controlled direct generator.

    Source: §6 · p. 7

  • Takeaways & Limitations

    The discrete evidence record supports structured content selection and auditable CBCT report generation; the system ranked second in automated and third in blinded clinical ODIN 2026 evaluation.

    Source: §6 · pp. 7–8

    The findings are limited by non-exhaustive report supervision, lack of independently adjudicated image-level truth, and a compact schema that underrepresents rare qualifiers and findings.

    Source: §6 · p. 8

Abstract

from arXiv · show

Dento-maxillofacial cone-beam CT (CBCT) reports may contain dozens of tooth-specific, anatomical, and spatial findings from a single 3D scan. Learning to generate such reports from limited clinical data is challenging because routine reports may not exhaustively document image findings, and a non-mention may reflect either absence or non-reporting. We present EviDent-CBCT, an evidence-bottlenecked framework designed for this incomplete supervision. An anatomy-aware network maps each CBCT scan to a discrete record of tooth-level, global, and tooth-IAC evidence. A dental-logic consistency projection reconciles incompatible evidence before a deterministic renderer and an image-blind local language model generate the report using only this record. For tooth-level evidence, reliability-aware training uses eligible non-mentions as reduced-weight negatives, while unreported global and tooth-IAC labels remain unknown. A metal-sensitive input channel preserves intensity cues from dental materials. Across three validation runs, EviDent-CBCT achieves $0.666\pm0.006$ merged evidence set-F1 and $0.402\pm0.003$ RadFact-Lite-Dental logical-F1, versus $0.371\pm0.018$ for the strongest controlled direct baseline. In the ODIN 2026 challenge, it ranked second in automated evaluation and third in blinded clinical Arena comparison on the hidden test set. These results support the discrete evidence record as an effective and auditable interface for CBCT report generation.

1 Introduction

CBCT report generation must recover and anatomically locate findings despite routine reports that may omit image findings. EviDent-CBCT separates evidence prediction from language realization through a discrete record used by an image-blind report generator.

1 Introduction

  • Routine reports document clinical practice rather than exhaustively annotating images, so a non-mention may reflect absence, omission, habitual non-reporting, or oversight.Aggregate non-mention patterns may offer negative signal without establishing absence in an individual case.Source: §1 · p. 1
  • EviDent-CBCT makes a consistency-projected discrete evidence record the sole case-specific input to an image-blind language model.Anatomy-aligned branches predict tooth-level, global, and tooth-IAC evidence; eligible tooth-level non-mentions receive reduced-weight negative supervision, while unreported global and tooth-IAC labels remain unknown.Source: §1 · p. 1, Figure 1 · p. 2
  • 0.402 ± 0.003 RadFact-Lite-Dental logical-F1 versus 0.371 ± 0.018 for the strongest controlled direct interface across three validation runs.EviDent-CBCT ranked second in automated evaluation and third in blinded clinical comparison in the ODIN 2026 challenge.Source: §1 · pp. 1–2
  • The framework is designed to separate visual evidence prediction from language realization and make selected report content auditable.The stated contributions include anatomy-aligned evidence learning under non-exhaustive supervision and controlled direct-generation comparisons across five visual interfaces, two report targets, and general-purpose and medically aligned decoders.Source: §1 · p. 2

2 Related Work

Prior medical report-generation approaches often condition decoders on continuous visual features, retrieved text, or soft diagnostic guidance. EviDent-CBCT requires report realization to proceed through an explicit discrete evidence record.

Medical report generation

  • Medical report-generation systems have progressed from recurrent and joint models to memory-, retrieval-, and region-guided architectures, with recent work connecting visual encoders to large language models.Most prior work targets chest radiographs, while CT2Rep extends direct generation to volumetric CT (Hamamci, Er, and Menze 2024).Source: §2 · pp. 2–3

Factuality and evidence interfaces

  • Prior factuality methods address coherence, entity, relation, or semantic agreement, grounding, preference optimization, and reasoning, but do not require an explicit record of predicted findings.Structured approaches including KERP, region-guided generation, and concept bottlenecks separate content prediction from verbalization (Li et al. 2019; Koh et al. 2020; Tanida et al. 2023).Source: §2 · p. 3

Dental CBCT Analysis

  • Dental CBCT learning has focused chiefly on anatomical segmentation, while Lumetti et al. formulate reporting as ontology-grounded structured prediction from clinician-validated targets (Lumetti et al. 2026).That system does not address narrative realization, and its task-specific schema and structured metrics are not directly comparable to free-text report generation.Source: §2 · p. 3

3 Problem Formulation

EviDent-CBCT replaces direct generation from a CBCT volume with report generation through a finite evidence record. Deterministic consistency correction and rendering precede an image-blind rewriter.

3 Problem Formulation

  • Direct generation estimates p(y | x), whereas EviDent-CBCT derives a finite evidence record from the volume and observed report.The record separates tooth-level findings indexed by FDI position, volume-level findings, and tooth-IAC and canal-course relations.Source: §3 · p. 3
  • Evidence prediction, dental-logic projection, deterministic rendering, and image-blind rewriting factorize report generation into distinct stages.Every draft content item comes from a tuple in the projected record; retaining the draft makes rewrite-induced deviations auditable and allows evidence prediction to be evaluated separately from realization.Source: §3 · p. 3

4 Method

EviDent-CBCT separates CBCT report generation into anatomy localization, evidence prediction, consistency correction, and language realization. Its discrete record remains the content interface between prediction and an image-blind rewriter.

Overview

  • The method isolates anatomical localization, report-derived evidence recovery, deterministic consistency correction, and textual realization through a discrete content interface.Stage 1 supplies anatomical masks and reusable structure tokens; Stage 2 produces a partially observed multilabel record; Stage 3 converts that record to prose.Source: Overview · p. 3

Reference Evidence Construction

  • Reference evidence is built from observed reports with source spans retained and a review-and-correction agent verifying structured items.The agent confirmed 98.3% on the first pass and corrected the remaining 1.7%; the resulting workflow structures observed reports rather than adjudicating them as exhaustive clinical truth.Source: Reference Evidence Construction · p. 3

Stage 1: Anatomy Segmentation and Feature Exposure

  • Stage 1 uses a 3D U-Mamba2 segmenter trained on ToothFairy to predict 46 anatomical classes and supply masks and structure tokens.It achieves macro Dice 0.899 over present structures and case-averaged Dice 0.880.Source: Stage 1: Anatomy Segmentation and… · p. 3

Stage 2: Reliability-Aware Evidence Prediction

  • Stage 2 aligns tooth, global, and tooth-IAC prediction branches with tooth-centered crops, a downsampled full volume, and posterior-mandible views.It combines segmentation context with appearance features and adds a raw-intensity channel to preserve variation among high-density dental materials and artifacts. Eligible tooth-level non-mentions receive reduced, class-dependent negative weights; unreported global and tooth-IAC labels remain unknown.Source: Stage 2: Reliability-Aware Evidence… · p. 4

Stage 3: Evidence-Bottlenecked Report Realization

  • Stage 3 removes incompatible predicted tuples with deterministic dental constraints, renders the corrected evidence into a canonical draft, and trains an image-blind rewriter to preserve its content.The local Qwen student receives only the deterministic draft during training and inference, with no CBCT, reference report, or parallel visual representation.Source: Stage 3: Evidence-Bottlenecked Report… · p. 4

5 Experiments

Experiments compare EviDent-CBCT with controlled direct generators and examine evidence prediction, report realization, rewrite fidelity, and hidden-test performance. Across three runs, the framework achieved .666 ± .006 merged Stage 2 logical set-F1 and .402 ± .003 final report logical-F1; it ranked second in ODIN 2026 automated evaluation and third in blinded clinical comparison.

Dataset

  • The paired ToothFairy4 subset contains 624 volumes and 1,001 reports; 561 volumes train and 63 validate, with conflicting binary labels excluded from supervision.All stages and direct baselines used the same volume-level split.Source: Dataset · pp. 4–5

Evaluation Metrics

  • Merged Stage 2 logical set-F1 measures exact structured tuple agreement before projection; RadFact-Lite-Dental evaluates bidirectional proposition entailment using fixed-reference main-paper metrics.Paired uncertainty was estimated by jointly bootstrapping matched cases and computing percentile 95% confidence intervals of paired differences.Source: Evaluation Metrics · p. 5

EviDent-CBCT vs. Controlled Direct Generation

  • Across three runs, .666 ± .006 merged Stage 2 logical set-F1 and .402 ± .003 final report logical-F1; rewritten output exceeded canonical Direct 2's .371 ± .018 F1.The paired gain over the strongest Med42 direct configuration was .0317 (95% bootstrap confidence interval [.0047, .0586]).Source: Table 1 · p. 6, EviDent-CBCT vs. Controlled Direct… · p. 6, Stage 2 Evidence Learning and Tooth… · p. 7
  • Canonical targets improved every decoder-interface pair; canonical Med42 Direct 2 reached .371 ± .018 F1, while Direct 3 and Direct 4 reached .362±.006 and .365±.007.The five interfaces use Qwen3-8B or Med42-8B; Direct 3 and Direct 4 retain continuous features, with Direct 4 adding auxiliary evidence supervision.Source: EviDent-CBCT vs. Controlled Direct… · p. 6

Stage-wise Analysis of Evidence-to-Report Pathway

  • Projecting richer evidence JSON into the compact schema and rendering it reduced logical-F1 from .957 to .589; Qwen realization raised it to .597.The consistency projection modified 178 tuple instances across 52 validation cases, raising template F1 from .389±.006 to .398±.004 and rewritten F1 from .398±.001 to .402±.003.Source: Stage-wise Analysis of… · pp. 6–7, Table 2 · p. 7

Auditability and Rewrite Fidelity

  • In 28/30 audited cases (93.3%), rewrites preserved all template propositions without polarity or anatomical reassignment; two substantive edits aligned more closely with references.The deterministic draft remains available for case-level inspection of rewrite changes.Source: Auditability and Rewrite Fidelity · p. 7

Stage 2 Evidence Learning and Tooth Ablations

  • Tooth-branch pu-weak supervision achieved .434 ± .028 precision, .538 ± .022 recall, and .480 ± .007 F1.Hard negatives raised F1 from .359 ± .018 to .467±.017; removing the metal-sensitive channel lowered F1 from .480 to .463, with other branches held fixed.Source: Stage 2 Evidence Learning and Tooth… · p. 7, Table 3 · p. 7

Official ODIN 2026 Hidden-Test Evaluation

  • On the ODIN 2026 hidden test set, EviDent-CBCT ranked second in automated evaluation and third in blinded clinical comparison.Its official aggregate score was 0.3033; the official RadFact protocol differs from RadFact-Lite-Dental, so absolute scores are not directly comparable.Source: Official ODIN 2026 Hidden-Test… · p. 7

6 Conclusion and Outlook

EviDent-CBCT separates evidence prediction from report realization and makes selected content auditable. It outperforms the strongest controlled direct generator on internal logical-F1 and places second in automated and third in blinded clinical ODIN 2026 evaluation, with limitations from incomplete supervision and schema coverage.

Conclusion

  • 0.402 ± 0.003 RadFact-Lite-Dental logical-F1, versus 0.371±0.018 for the strongest controlled direct generator, across three validation runs.The discrete record enables deterministic dental-consistency correction and makes selected report content auditable.Source: §6 · p. 7
  • Second in ODIN 2026 automated evaluation and third in blinded clinical comparison on the hidden test set.The official RadFact protocol differs from internal RadFact-Lite-Dental evaluation; absolute scores are not directly comparable.Source: §6 · p. 8, Table 4 · p. 8

Outlook

  • The findings are limited by non-exhaustive supervision, no independently adjudicated image-level truth, and a compact schema that underrepresents rare findings and qualifiers.Hidden-test references are unavailable for detailed error analysis.Source: §6 · p. 8

Supplementary Material of EviDent-CBCT

The supplement clarifies that reports provide incomplete, non-adjudicated supervision, and distinguishes the rich report-derived evidence from the compact prediction schema. Its audits locate substantial information loss in schema projection and report realization rather than initial source-span extraction.

A Scope of the Report-Derived Supervision

  • Non-mentions cannot be distinguished as absence, omission, habitual non-reporting, oversight, or reader disagreement from reports alone; explicit inaccuracies are possible but unmeasured.Eligible tooth-level non-mentions receive one of three policies, while unreported global and tooth-IAC labels remain unknown; the resulting metric measures agreement with reports, not external clinical accuracy.Source: A Scope of the Report-Derived… · p. 10

Construction and independent verification

  • A GPT-5.5 extraction pass attaches supporting report text to structured evidence, then a Claude Opus 4.8 reviewer checks support and anatomical assignments and requests corrections when needed.The 98.3% first-pass confirmation rate measures reviewer acceptance, not dentist validation or correctness against an independently read CBCT volume.Source: Construction and independent… · p. 10

Full JSON versus the prediction schema

  • The full evidence JSON retains free-form facts and source text, whereas Stage 2 projects recurrent, learnable content into 11 tooth, 12 global, and 8 relation categories.Rare details can remain auditable in the JSON without becoming dedicated image-prediction classes.Source: Full JSON versus the prediction schema · p. 10

Complete Stage 2 evidence schema

  • Positive counts are frequencies in observed reports, not clinical prevalence; unreported sparse global and relation labels remain unknown, while eligible tooth non-mentions receive reliability weights.Explicit positive and negative statements receive weight 1; unavailable positions and contradictory statements receive weight 0.Source: Complete Stage 2 evidence schema · p. 10

Coverage protocol

  • The coverage audit reassembles 555 unique report clues from attached text spans, retaining each as an exact substring after whitespace normalization.Literal coverage measures retained wording, while semantic coverage measures whether selected spans support propositions recovered from the report.Source: Coverage protocol · p. 10
  • 96.5% mean literal coverage and 0.957 logical-F1 show that retained source spans preserve substantial report content; residuals concentrate in low-frequency metadata, classifications, and broad normal statements.The compact fields emphasize macroscopic, recurrent concepts, and residual concepts are too sparse in this corpus for separate supervised classes.Source: Coverage protocol · p. 10, Residual content · p. 10

Coverage-to-ceiling interpretation

  • 0.957 logical-F1 for reassembled source clues falls to 0.589 after projection into the compact schema and deterministic rendering, then to 0.398 ± 0.001 with predicted evidence and rewriting.Rewriting reference-derived evidence raises logical-F1 from 0.589 to 0.597, an operational ceiling for the compact schema-to-report pathway.Source: Coverage-to-ceiling interpretation · pp. 10–11

C Implementation Details

Implementation uses reliability-weighted tooth-level supervision, anatomy-conditioned branches, and fixed crop procedures to construct evidence. A deterministic consistency projection precedes rendering, and the language model receives only the resulting draft.

Tooth-level supervision

  • Eligible negatives are sampled without replacement toward a negative-to-positive ratio of 0.5 ≤ n−/n+ < 2 when enough negatives are available, while predefined class-dependent reliability weights remain in the loss.The weights reflect how reliably non-mention implies absence; for example, non-mention of caries receives lower weight than non-mention of missingness for a mask-present tooth.Source: C Implementation Details · p. 11

Tooth branch

  • Tooth-mask centers and principal axes define oriented crops; mask-absent positions receive interpolated fixed crops without tooth-branch supervision, and their missing status comes from the Stage 1 mask.For classifiable teeth, predicted missingness conditions the remaining attributes through a detached feature pathway.Source: C Implementation Details · p. 11

Input crops

  • Tooth crops enclose the full segmented tooth and expand it by 6 mm in every direction, while relation crops use a fixed 80 × 96 × 64-mm field of view resized to 80 × 128 × 128.Tooth-dependent effective voxel spacing follows from resizing differently sized teeth to a fixed input size.Source: C Implementation Details · p. 12

Evidence projection and realization

  • The dental-logic projection removes incompatible tuples, omits certain IAC relations for missing teeth, and resolves competing canal-course tuples by predicted confidence before rendering.The projection introduces no positive finding; the Qwen rewriter receives only the deterministic draft.Source: C Implementation Details · p. 12

D Rewrite Content-Preservation Audit

The rewrite preserved every template proposition in 28 of 30 audited cases, but was not strictly lossless. Across controlled comparisons, EviDent-CBCT’s final reports outperformed direct generation, with the advantage persisting under a multiple-reference protocol.

D Rewrite Content-Preservation Audit

  • The rewrite preserved all template propositions in 28 of 30 audited cases (93.3%), but two substantive departures changed predicted findings.Both departures moved toward the reference report; the deterministic template remains an auditable fallback.Source: D Rewrite Content-Preservation Audit · p. 12, Table 10 · p. 13

Controlled interfaces and supervision

  • Direct 5 freezes Stage 2 branches while retaining continuous features, separating representation preservation from exposing a discrete evidence record to the generator.Direct 4 instead adds evidence-prediction heads to Direct 3; its ground-truth evidence forms training loss only and is not supplied at inference.Source: Controlled interfaces and supervision · pp. 12–13

Run-level results

  • Canonical targets improved all ten decoder-interface pairs; Med42 raised the strongest direct result from .331 with Qwen3 to .371.Direct 5 reached .327 with Qwen3 and .360 with Med42, below trainable Direct 3 for both decoders.Source: Run-level results · p. 13

Case-level paired bootstrap

  • Canonical supervision improves every decoder-interface pair, while the final consistency-projected comparison favors EviDent-CBCT over direct generation.The paired bootstrap jointly resampled matched cases 100,000 times and compared against Med42 Direct 2 with canonical supervision.Source: Run-level results · p. 13, Case-level paired bootstrap · p. 13
  • The final consistency-projected outputs scored .4023 logical-F1 versus .3706 for direct generation, a paired difference of .0317 (95% CI [.0047, .0586]).The interval measures case-sampling uncertainty after averaging model runs, not uncertainty over training seeds.Source: Case-level paired bootstrap · p. 13

Sensitivity to multiple reports per volume

  • Under mean-over-references evaluation, both EviDent-CBCT realizations scored .374 logical-F1 versus .333 for the strongest direct generator.The locally rewritten output’s paired advantage was .0402 (95% CI [.0170, .0637]); its difference from the template was negligible at −.0004 (95% CI [−.0074, .0065]).Source: Sensitivity to multiple reports per… · p. 13

F Global and Relation Fusion Sweep

A single-run architecture sweep compared global and relation interfaces using merged Stage 2 logical set-F1. Its narrow score range did not establish a generally superior fusion choice.

F Global and Relation Fusion Sweep

G Evaluation Semantics

The paper distinguishes an LLM-free structured-agreement measure from report-level RadFact-Lite-Dental evaluation. Its dental evaluator more closely matches author-audited case scores than the original evaluator, despite a small aggregate offset.

Merged Stage 2 logical set-F1

  • Merged Stage 2 logical set-F1 uses exact structured tuple matches pooled across tooth, global, and relation branches, and is not the report-level RadFact score.It computes corpus-level precision and recall from case-level values and takes their harmonic mean.Source: Merged Stage 2 logical set-F1 · p. 14

RadFact-Lite-Dental

  • RadFact-Lite-Dental adapts bidirectional proposition entailment to dental reports, preserving qualifiers and checking tooth position, anatomy, side, polarity, and finding type.The validation protocol fixes the lexicographically smallest report identifier as each volume’s reference; ODIN hidden-test RadFact uses a different protocol and is not numerically comparable.Source: RadFact-Lite-Dental · p. 15
  • The dental evaluator’s case-level MAE was .043 and Pearson correlation .881, while its 1,375 entailment decisions reached .949 accuracy against the author audit.The original evaluator’s case-level MAE was .091 despite matching the audited cohort F1 at .410; its errors largely canceled in aggregate.Source: RadFact-Lite-Dental · p. 15

H Class-wise Stage 2 Branch Analysis

Among tooth-level non-mention policies, pu-weak achieved the highest tooth logical set-F1, while some rare-label rankings remain descriptive. Qualitative cases show differences in recovered tooth findings and canal relationships, alongside errors by both systems.

H Class-wise Stage 2 Branch Analysis

  • Pu-weak produced the highest tooth logical set-F1, followed by hard and explicit; it had the highest class F1 for six of eleven categories.Its advantage was not uniform across rare labels, whose small validation supports make policy-specific ordering descriptive.Source: H Class-wise Stage 2 Branch Analysis · p. 15, Table 15 · p. 16
  • The global branch reached .884 branch logical set-F1, while the relation branch reached .537 over the complete predicted relation set.Perfect scores for periodontitis and poor-quality scans had limited positive support and no observed negatives; buccal-course relation categories remained precision-limited.Source: H Class-wise Stage 2 Branch Analysis · p. 15, Table 16 · p. 16

I Qualitative Report Comparisons

  • In the treatment-rich and edentulous examples, EviDent-CBCT recovered broader tooth findings and all mandibular missing teeth, while both systems made errors.The direct model omitted or misclassified more tooth-level findings in the edentulous case; both models struggled with maxillary root remnants.Source: Figure 3 · p. 17, Figure 4 · p. 18
  • For bilateral impactions, EviDent-CBCT captured both teeth and canal relationships in one case and all core propositions in another, while the direct model missed canal relations.The examples also show omissions and unsupported findings by EviDent-CBCT; color annotations distinguish matches, partial matches, unsupported predictions, and unrecovered reference facts.Source: I Qualitative Report Comparisons · p. 15, Figure 5 · p. 19, Figure 6 · p. 20
Loading 2610.02375v1…