Source-linked AI summary
Entity-Constrained CBCT Retrieval for Low-Resource Dental Record Completion
Nhi Ngoc-Yen Nguyen, Thai Nguyen, Kiet Huynh Cao Tuan, Huy-Hieu Pham
TL;DR
Low-resource CBCT record completion is difficult because seven clinical fields rely on different evidence and unsupported FDI or ICD entities can be penalized. ECCR combines a complete-record text prior with frozen 3D retrieval, allowing only entity-safe Diagnosis additions. It improves public-validation performance over full-record retrieval and a static prior, while ranking second on the final test evaluation.
Problem
Low-resource CBCT record completion requires seven-field records from 50 labeled cases, where unsupported FDI positions or ICD codes can incur penalties.
Method
ECCR uses a corpus-derived prior for all fields and appends frozen-encoder Diagnosis evidence only when it introduces no new FDI positions or ICD codes.
Results
0.3134: ECCR’s public-validation weighted score exceeded full-record multimodal retrieval at 0.2237 and the static text-only prior at 0.2915; it obtained 11.37 and ranked second on final evaluation.
Takeaways & Limitations
In extreme low-resource settings, controlling which multimodal evidence may modify a record can be more reliable than transferring an entire retrieved record.
Takeaways & Limitations
The guarantee covers only FDI positions and dental ICD-10 codes, while prior coverage caps entity recall and Diagnosis-only authority forfeits the entire Oral exam aspect.
Abstract
from arXiv · showhide
Completing dental records from cone-beam computed tomography (CBCT) is difficult when annotation is scarce and individual clinical fields are supported by different types of evidence. MMDental Task 3 requires seven-field record completion from only 50 labeled CBCT cases and scores the correctness of structured FDI positions and ICD codes; consequently, a visually plausible retrieved record can still be harmful when it introduces an unsupported entity. We propose Entity-Constrained CBCT-Guided Retrieval (ECCR), a parameter-free framework that separates evidence availability from evidence authority. A corpus-derived prior first supplies the complete record. A frozen 3D encoder retrieves image-conditioned Diagnosis evidence, which is appended only if it does not expand the prior FDI or ICD entity set, so the asserted entity set is invariant by construction. On public validation, ECCR reaches a weighted score of 0.3134, improving on both full-record multimodal retrieval (0.2237) and a static text-only prior (0.2915); the guard blocks 63.3% of retrieved candidates, each of which would otherwise have injected an FDI position or ICD code absent from the prior. On the final test evaluation, ECCR obtains 11.37 of a 97.4-point attainable maximum, securing second place overall. The result indicates that, in an extreme low-resource setting, controlling what multimodal evidence is allowed to modify can be more reliable than transferring an entire retrieved record.
1 Introduction
ECCR addresses low-resource CBCT record completion by separating retrieval from field-level authority, because unsupported entities can harm strictly evaluated dental records. It uses a prior for complete records and permits only entity-safe Diagnosis updates.
- MMDental Task 3 requires seven-field records from only 50 labeled CBCT cases, with penalties for incorrect FDI tooth positions and ICD codes.
- 0.2237: unrestricted M3D-CLIP neighbor replacement underperformed the text-only BLEU-medoid prior at 0.2915 on public validation.
- ECCR supplies all seven fields from a corpus-derived prior and restricts retrieved updates to Diagnosis when no new FDI positions or ICD codes are introduced.
- The framework formulates low-resource record completion as an evidence-authority problem, distinguishing retrieval capability from permissions to modify fields.
2 Related Work
Prior work makes image-conditioned medical evidence accessible and improves retrieval, but does not establish which clinical fields may be safely modified under extreme data scarcity. ECCR uses deterministic structured-entity checks instead of training a dedicated factuality model.
- Large 3D vision–language models connect volumetric images with medical text, while report-generation systems retrieve or synthesize radiological reports.
- Retrieval-augmented generation improves access to relevant evidence but risks introducing unsupported clinical findings.
- With only 50 cases, ECCR uses extracted FDI positions and ICD codes as a parameter-free, deterministic admission rule rather than training a dedicated factuality model.
3 Dataset Analysis and Design Motivation
The dataset’s heterogeneous fields, concentrated structured entities, and uniform volume geometry motivate a corpus prior plus optional, rather than universal, image-conditioned updates.
- 161 labeled visits from 50 cases are aggregated by filename into one structured record per CBCT volume.
- Table 1 reports field heterogeneity using filled-row percentages and distinct non-empty string counts across the labeled visits.
- Doctor advices is strongly templated, whereas Handle and Oral check are nearly visit-specific and several history fields are frequently absent.
- 759 tooth mentions cover 45 FDI positions and 116 ICD mentions cover 29 codes, with both entity types concentrated in a small set of frequent values.
- All 50 labeled volumes share 640 × 640 × 400 geometry at isotropic 0.25 mm spacing, so geometry does not explain variation in this split.
4 Method
ECCR combines a corpus-derived text prior with frozen 3D retrieval, then governs whether evidence may modify each field. Only Diagnosis can be updated, and admission preserves the prior’s detected entity set.
- 4.1 Problem Formulation: The method separates evidence availability from evidence authority using a training-derived prior, retrieved text, and a binary field-admission function.
- 4.1 Problem Formulation: Full authority replaces the record, absent authority retains the prior, and bounded authority restricts admission by field and content.
- 4.3 Entity-Constrained Diagnosis Update: The prior supplies all seven fields, while retrieval may append only Diagnosis evidence; non-Diagnosis fields have Af = 0.
- 4.2 Text Prior and Image Retrieval: A frozen Med3DVLM VisionTower retrieves top-1 neighbors from cached, normalized bank features without task-specific training.
- 4.3 Entity-Constrained Diagnosis Update: Admission extracts FDI positions and dental ICD-10 codes and accepts a candidate only when its entity set is contained in the prior’s set.
- 4.3 Entity-Constrained Diagnosis Update: Retrieval cannot alter the detected entity set: the prior bounds entity precision, while its coverage caps entity recall.
5 Experiments
The experiments compare evidence-authority policies using an official public-validation score, an offline LOO proxy, admission traces, and final leaderboard results. ECCR improves validation performance and ranks second overall, while its analyses expose proxy limitations and field-specific costs.
- 5.1 Protocol and Implementation: 0.3134 is ECCR’s public-validation weighted score, exceeding full-record retrieval at 0.2237 and the text-only prior at 0.2915.Table 3 compares full authority, no authority, and entity-compatible Diagnosis updates.
- 5.2 Evidence-Authority Comparison and Update Trace: 0.0219 is the gain from entity-compatible Diagnosis evidence over the text-only prior, reaching 0.3134 and second place on the public leaderboard.The comparison supports selective evidence use but does not isolate authority as the sole cause of the full-record gap because encoders differ.
- 5.2 Evidence-Authority Comparison and Update Trace: 40 of 50 LOO folds contain retrieved FDI or ICD entities absent from the prior, and Equation (4) guarantees zero injections.Replacement and naive appending propagate these entities, whereas entity-safe appending blocks them; replacement costs 0.0289 and naive versus safe appending differs by 0.0002 on the proxy.
- 5.2 Evidence-Authority Comparison and Update Trace: 63.3% of 49 public-validation retrievals are blocked by the constraint, with 31 candidates blocked and 18 admitted.The admission trace is separate from the 50-fold LOO analysis.
- 5.3 Reconciling the Offline Proxy with the Official Score: The offline proxy cannot penalize unsupported entities because smoothed BLEU and ROUGE-L measure lexical overlap with a single reference.Therefore, proxy parity between appending rules does not establish the safety property, which Equation (4) establishes by construction.
- 5.4 Final Leaderboard Result and Score Decomposition: A static medoid prior caps tooth-level Diagnosis F1 at 0.016, while Oral exam scores zero because the admission rule is restricted to Diagnosis.Oral check is near visit-specific, with 146 unique values in 161 visits; Past history also scores zero because aggregation leaves it empty.
- 5.4 Final Leaderboard Result and Score Decomposition: 11.37 is ECCR’s final score against a 97.4-point attainable maximum, ranking second overall.ECCR remains ahead of the third-placed system because Record completeness contributes 3.76 versus 1.68 despite a lower Diagnosis sub-score.
- 5.5 Admission Behaviour: Conflicting candidates occur throughout the similarity range, so cosine similarity cannot substitute for entity-aware verification.The analysis separates retrieval capability and evidence authority while identifying spatial localization and prior coverage as future representation constraints.
6 Conclusion and Limitations
ECCR bounds multimodal updates to preserve the prior’s asserted FDI and ICD entity set, while its guarantees and comparisons remain limited in scope.
- Conclusion: ECCR preserves the asserted FDI and dental ICD-10 entity set by constraining candidate Diagnosis updates.The guarantee does not extend to free-text clinical concepts.
- Limitations: The invariant bounds entity precision by the prior but caps entity recall by the prior’s coverage.This limitation leaves tooth-level Diagnosis F1 at 0.016.
- Limitations: Restricting authority to Diagnosis forfeits the entire Oral exam aspect, worth 20% of the weight.A per-field guard for oral findings is identified as the most valuable extension.
- Limitations: The comparison against full-record retrieval confounds authority with encoder choice, and the fixed single split provides no variance estimate.Future work includes per-field admission rules and field-specific entity vocabularies.