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

cs.CVOct 1, 2026

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

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

Organizations: King’s College London

Abstract

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.

Figures & tables

Figure 1: Overview of EviDent-CBCT. During training, routine clinical reports are converted into source-linked reference evidence that preserves verbatim supporting spans; explicit labels supervise all branches; eligible tooth-level non-mentions provide reduced-weight negative supervision. At inference, Stage 1 extracts anatomical masks and structure tokens from the CBCT volume. Stage 2 combines tooth-centered, whole-volume, and posterior-mandibular views to predict tooth, global, and tooth-IAC findings, which are consolidated into a discrete evidence record. A dental-logic consistency projection reconciles incompatible evidence before Stage 3 deterministically renders a canonical draft, which serves as the sole case-specific input to an image-blind local Qwen rewriter. The rewriter is trained through GPT-based content-preserving distillation on drafts constructed from consistency-projected predicted evidence. Solid arrows indicate inference; orange arrows signify training-only supervision; and light-blue arrows denote reduced-weight supervision from eligible tooth-level non-mentions.

Figure 2: Comparison of the inference-time interfaces used by the controlled direct generators and EviDent-CBCT. Direct 1 and Direct 2 condition a language decoder on frozen or fine-tuned Stage 1 representations. Direct 3 fine-tunes Stage 2-checkpoint-initialized branches and supplies their continuous features to the decoder; Direct 4 adds training-only auxiliary evidence supervision, while Direct 5 freezes the Stage 2 branches. These five interfaces are instantiated with Qwen3-8B and Med42-8B decoders. EviDent-CBCT instead supplies a consistency-projected evidence record as the only case-specific input to an image-blind realizer. Its deterministic draft exposes the predicted findings and makes rewrite-induced deviations auditable.

Family	Visual/report interface	Report target	Logical P	Logical R	Logical F1	BLEU-4	METEOR
Direct Qwen3-8B	Direct 1: frozen Stage 1 tokens	Raw	$.216\pm.016$	$.245\pm.036$	$.229\pm.018$	$.121\pm.005$	$.352\pm.004$
	Direct 1: frozen Stage 1 tokens	Canonical	$.295\pm.033$	$.294\pm.043$	$.294\pm.034$	$.095\pm.004$	$.349\pm.006$
	Direct 2: fine-tuned Stage 1 encoder	Raw	$.257\pm.027$	$.251\pm.009$	$.253\pm.011$	$.135\pm.007$	$.338\pm.012$
	Direct 2: fine-tuned Stage 1 encoder	Canonical	$.331\pm.056$	$.315\pm.023$	$.322\pm.036$	$.093\pm.007$	$.355\pm.005$
	Direct 3: continuous Stage 2 features	Raw	$.238\pm.034$	$.236\pm.030$	$.236\pm.025$	$.124\pm.015$	$.325\pm.024$
	Direct 3: continuous Stage 2 features	Canonical	$.351\pm.004$	$.314\pm.009$	$.331\pm.005$	$.099\pm.003$	$.356\pm.005$

Table 1: Validation set results, reported as mean

\pm

std over three runs. Logical metrics use RadFact-Lite-Dental.

Candidate report	Logical F1
Reference report vs. itself	1.000
Verbatim source clues from evidence extraction	.957
Reference-derived evidence $\rightarrow$ template	.589
Reference-derived evidence $\rightarrow$ template $\rightarrow$ Qwen	.597
Predicted evidence $\rightarrow$ template	$.389\pm.006$
Predicted evidence $\rightarrow$ template $\rightarrow$ Qwen	$.398\pm.001$

Table 2: Evidence-to-report pathway decomposition and reference-derived operational ceiling. Reference-derived rows are evaluated once; predicted-evidence rows are reported as mean

\pm

std over three runs.

Axis	Variant	Logical P	Logical R	Logical F1
Base	gated, metal, pu-weak	$.434\pm.028$	$.538\pm.022$	$\mathbf{.480\pm.007}$
Supervision	hard	$.408\pm.017$	$\mathbf{.548\pm.030}$	$.467\pm.017$
	explicit	$.405\pm.057$	$.328\pm.035$	$.359\pm.018$
Fusion	concat	$\mathbf{.435\pm.036}$	$.525\pm.033$	$.474\pm.015$
	visual-only	$.422\pm.018$	$.534\pm.033$	$.471\pm.004$
	scratch-only	$.379\pm.021$	$.512\pm.039$	$.435\pm.001$

Table 3: Tooth-branch Stage 2 logical set metrics, reported as mean

\pm

std over three model runs. All non-ablated branches and inputs are held fixed.

Team	RadFact F1	BLEU-4	METEOR	Aggregate Score	Arena Score
shivam.singh	0.3924	0.1476	0.3269	0.3614	1307
EviDent-CBCT(ours)	0.3245	0.1206	0.3171	0.3033	985
ViktorB	0.2321	0.0865	0.2828	0.2226	1172

Table 4: Official ODIN 2026 ToothFairy4 hidden-test results for the top three teams. The official RadFact protocol differs from RadFact-Lite-Dental used in our internal validation.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Schema field	Meaning	Total $n_{+}$	Observed $w_{-}$	Report-level $w_{-}$
missing	Tooth absent at the indexed FDI position	924	.95	.95
implant	Endosseous implant at the indexed position	256	.80	.50
crown	Prosthetic crown at the indexed position	695	.60	.40
restoration	Coronal restoration or filling, distinct from a crown	1,792	.50	.30
endodontic_treatment	Root-canal or other endodontic treatment	766	.70	.40
prosthesis_related	Bridge, pontic, or other prosthetic involvement	1,305	.50	.30

Appendix

Table 5: Complete tooth-level prediction schema, positive sample counts, and class-dependent pu-weak weights. Observed denotes a non-mention for a tooth observed elsewhere in the report; report-level denotes an otherwise eligible non-mention inferred only from report silence.

Schema field	Meaning	Total $n_{+}$
maxilla_partially_included	Maxilla only partly included in the scan field	236
maxilla_not_included	Maxilla excluded from the scan field	234
maxillary_sinuses_not_included	Maxillary sinuses excluded from the scan field	173
mandibular_condyles_not_included	Mandibular condyles excluded from the scan field	588
periodontitis_present	Reported periodontal disease	207
atrophy_maxilla	Maxillary alveolar atrophy	53

Appendix

Table 6: Complete global prediction schema and positive sample counts.

Schema field	Meaning	Total $n_{+}$
38_close_to_left_IAC	Tooth 38 close to the left IAC	168
48_close_to_right_IAC	Tooth 48 close to the right IAC	168
any_tooth_close_to_IAC	Any reported tooth-IAC proximity	219
lesion_close_to_IAC	Lesion close to an IAC	14
left_IAC_predominantly_lingual	Predominantly lingual course of the left IAC	236
left_IAC_predominantly_buccal	Predominantly buccal course of the left IAC	60

Appendix

Table 7: Complete relation prediction schema and positive sample counts. IAC denotes the inferior alveolar canal.

Residual category	Cases
Image orientation, laterality discrepancy, or acquisition note	3
Low-frequency dental classification (Kennedy class)	1
Explicit complete-dentition or complete-arch statement	2
Recurring omission of common dental findings	0

Appendix

Table 8: Substantive report content not fully represented by retained evidence spans. Counts denote affected evaluation cases; categories may overlap within a case.

Branch	View and tensor size	Resolution/channels	Encoder and fusion
Global	Whole available scan; $96\times 192\times 192$	Case-dependent effective spacing; 1 normalized CT + 5 semantic channels	Seven-stage 3D residual encoder with widths $[32,64,128,256,320,320,320]$ , block counts $[1,3,4,6,6,6,6]$ , and strides $[1,2,2,2,2,2,1]$ ; terminal 320-channel MambaLayer2, global average pooling, and a 12-label MLP. This branch is trained scratch-only.
Tooth	One PCA-aligned crop per FDI position; $64\times 64\times 96$ , with the tooth long axis mapped to $x$	Tooth-dependent spacing after adding 6 mm on each crop axis; 1 anatomical CT + 1 metal-sensitive CT + 8 semantic channels	SmallCropEncoder: three $3\times 3\times 3$ convolutions with 12, 24, and 48 channels, instance normalization and LeakyReLU, pooling after the first two blocks, global pooling, and a 192-D projection. The crop feature is gated with a 192-D projection of the 1,408-D Stage 1 tooth token and a 16-D ten-region embedding, yielding a 256-D feature for label-specific heads; the missing head conditions the remaining attributes.
Relation	Paired left/right posterior-mandibular crops; each $80\times 128\times 128$ from an $80\times 96\times 64$ -mm FOV	$1.0\times 0.75\times 0.5$ mm; 1 normalized CT + 5 semantic channels	The same 12-24-48 SmallCropEncoder produces a 192-D feature per side. Left/right features and 16-D side embeddings are combined with projected Stage 1 relation tokens by gated fusion into 384 dimensions, followed by an eight-label MLP.

Appendix

Table 9: Stage 2 input geometry and encoder architecture in the primary configuration. CT and semantic-mask channels are concatenated before the branch encoder.

Audit outcome	Cases
All template propositions preserved	28
Any substantive content change	2
Omission	1
Finding-type substitution	1
Polarity reversal	0
FDI reassignment	0

Appendix

Table 10: Author audit of content changes introduced by Qwen rewriting. Categories can overlap within a case.

Family	Visual interface	Auxiliary evidence	Target	Run 1	Run 2	Run 3
Direct Qwen3-8B	Direct 1: frozen Stage 1	No	Raw	.235	.242	.209
	Direct 1: frozen Stage 1	No	Canonical	.325	.258	.297
	Direct 2: fine-tuned Stage 1	No	Raw	.244	.265	.251
	Direct 2: fine-tuned Stage 1	No	Canonical	.283	.355	.327
	Direct 3: continuous Stage 2 features	No	Raw	.240	.258	.210
	Direct 3: continuous Stage 2 features	No	Canonical	.328	.330	.337

Appendix

Table 11: Run-level RadFact-Lite-Dental logical-F1 for the controlled report-generation comparison.

Reference protocol	Candidate output	Logical P	Logical R	Logical F1
Fixed lexicographic	Med42 Direct 2, canonical (direct best)	$.388\pm.019$	$.354\pm.017$	$.371\pm.018$
	EviDent-CBCT template	$.372\pm.002$	$.428\pm.008$	$.398\pm.004$
	EviDent-CBCT local rewrite	$.384\pm.002$	$.423\pm.004$	$.402\pm.003$
Mean over all references	Med42 Direct 2, canonical (direct best)	$.361\pm.013$	$.310\pm.013$	$.333\pm.013$
	EviDent-CBCT template	$.369\pm.007$	$.379\pm.012$	$.374\pm.009$
	EviDent-CBCT local rewrite	$.374\pm.003$	$.373\pm.011$	$.374\pm.006$

Appendix

Table 12: Sensitivity of the final comparison to the reference protocol, reported as mean

\pm

std over three model runs. Under mean-over-references evaluation, logical precision and recall are averaged within each volume before corpus aggregation.

Global	Relation	Logical P	Logical R	Logical F1
hybrid	hybrid	.653	.692	.672
hybrid	scratch	.651	.691	.670
hybrid	visual	.651	.695	.672
scratch	hybrid	.652	.696	.673
scratch	visual	.650	.699	.674
visual	hybrid	.655	.702	.678

Appendix

Table 13: Global/relation interface sweep using merged Stage 2 exact-tuple agreement. The narrow range does not support a claim that one fusion choice is intrinsically superior.

Evaluation	Logical P	Logical R	Logical F1	F1 MAE	Pearson $r$
Author-audited gold	.364	.468	.410	–	–
RadFact-Lite v0.1.0	.371	.458	.410	.091	.587
RadFact-Lite-Dental	.366	.444	.401	.043	.881

Appendix

Table 14: Agreement with author-audited gold scores on 30 stratified report pairs. MAE and Pearson correlation are computed over case-level logical-F1.

	pu-weak (base)			explicit			hard
Finding	Logical P	Logical R	Logical F1	Logical P	Logical R	Logical F1	Logical P	Logical R	Logical F1
Missing tooth	$.412\pm.021$	$.622\pm.087$	$\mathbf{.494\pm.025}$	$.346\pm.027$	$.524\pm.075$	$.413\pm.016$	$.368\pm.024$	$.694\pm.084$	$.478\pm.002$
Implant	$.325\pm.013$	$.617\pm.021$	$.426\pm.014$	$.432\pm.092$	$.543\pm.057$	$\mathbf{.473\pm.046}$	$.338\pm.037$	$.580\pm.057$	$.426\pm.035$
Crown	$.452\pm.013$	$.646\pm.065$	$\mathbf{.530\pm.012}$	$.233\pm.101$	$.646\pm.053$	$.331\pm.111$	$.427\pm.031$	$.663\pm.079$	$.516\pm.005$
Restoration	$.360\pm.018$	$.665\pm.026$	$.466\pm.011$	$.161\pm.002$	$.943\pm.010$	$.275\pm.004$	$.424\pm.030$	$.542\pm.048$	$\mathbf{.474\pm.016}$
Endodontic treatment	$.413\pm.036$	$.595\pm.097$	$\mathbf{.482\pm.010}$	$.385\pm.083$	$.356\pm.035$	$.367\pm.049$	$.382\pm.042$	$.595\pm.043$	$.464\pm.035$

Appendix

Table 15: Class-wise tooth-branch Stage 2 performance under the three non-mention policies, reported as mean

\pm

std over three model runs. The final row is the case-grouped branch logical set score over all tooth categories.

Branch	Finding	Known	Positive	Logical P	Logical R	Logical F1
Global	Partial maxilla coverage	54	27	.579	.815	.677
	Maxilla excluded	57	21	.654	.810	.723
	Maxillary sinuses excluded	30	14	.933	1.000	.966
	Mandibular condyles excluded	59	58	.983	1.000	.991
	Periodontitis	19	19	1.000	1.000	1.000
	Maxillary atrophy	8	6	1.000	.833	.909

Appendix

Table 16: Class-wise Stage 2 performance of the base global and relation branches. Counts expose the support available for each category; branch logical-set rows are case-grouped and are not averages of the class rows.

Figure 3: A treatment-rich case with partial maxillary coverage. EviDent-CBCT recovers broader reference-supported tooth-status, restorative, and endodontic findings than the best direct model, although both predictions contain unsupported or mismatched details. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact.

Figure 4: A heavily edentulous case with a complex treatment history. EviDent-CBCT recovers the complete set of mandibular missing teeth and correctly localizes several restorative and endodontic findings. The direct model omits or misclassifies more tooth-level findings, while both models remain challenged by maxillary root remnants. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact.

Figure 5: A mandibular-only case with bilateral complete bony impaction of teeth 38 and 48. EviDent-CBCT captures both impacted teeth and their relationships with the mandibular canal, but omits the complete-bony qualifier and introduces several unsupported findings. The direct model misclassifies tooth 48 as absent and adds unsupported tooth-level abnormalities. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact.

Figure 6: A case with bilateral canal–impacted-third-molar relationships and partial maxillary inclusion. EviDent-CBCT covers all core reference propositions, including both canal relationships, the negative osseous finding, and the acquisition coverage. The direct model identifies the impactions but omits their canal relationships and incorrectly reports that the maxilla is not included. In each composite, predictions appear above their annotated references. Dark green indicates a match, light green a partial or underspecified match, blue an unsupported or contradictory prediction, and red an unrecovered reference fact.

Explore similar work

Entity-Constrained CBCT Retrieval for Low-Resource Dental Record Completion
Aug 22, 2026Nhi Ngoc-Yen Nguyen, Thai Nguyen, Kiet Huynh Cao Tuan +1Evidence Retrieval Information Retrieval
VinUni-Illinois Smart Health Center, VinUniversity, Hanoi, Vietnam · Vietnam National University Ho Chi Minh City, University of Science, Vietnam
Clinical Reasoning Under a Partially Observed Objective in Cone Beam CT Report Generation
Sep 2, 2026Ajo Babu George, Govind Arun, Sidharth N Krishna +1Radiology Report Generation Radiology Report Evaluation
DiceMed, Cuttack, Odisha, India · University of Maryland, College Park, Maryland, USA · Indira Gandhi National Open University, New Delhi, India +1
X-Splat: Gaussian Splatting for 3D CBCT Generation from Single Panoramic Radiograph
Jul 2, 2026Tomasz Szczepański, Szymon Płotka, Michal K. Grzeszczyk +23D Gaussian Splatting Cone-Beam CT
Sano Centre for Computational Medicine, Poland · Jagiellonian University, Poland · Warsaw University of Technology, Poland +2

Aug 22, 2026, cs.CVJ/K move · Enter open · S save

Entity-Constrained CBCT Retrieval for Low-Resource Dental Record Completion

Nhi Ngoc-Yen Nguyen, Thai Nguyen, Kiet Huynh Cao Tuan, Huy-Hieu Pham

VinUni-Illinois Smart Health Center, VinUniversity, Hanoi, Vietnam · Vietnam National University Ho Chi Minh City, University of Science, Vietnam

Evidence Retrieval Information Retrieval Cone-Beam CT

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Entity-Constrained CBCT Retrieval for Low-Resource Dental Record Completion

Clinical Reasoning Under a Partially Observed Objective in Cone Beam CT Report Generation

X-Splat: Gaussian Splatting for 3D CBCT Generation from Single Panoramic Radiograph