EviDent-CBCT
An evidence bottleneck framework for automated dental CBCT imaging report generation. A discrete evidence recording layer decouples imaging data from textual reports, supporting auditable and traceable clinical report generation under non-exhaustive supervision scenarios.
Timeline (last 90 days)
EviDent-CBCT is an Evidence-Bottlenecked framework for dental CBCT report generation, specifically designed for incomplete annotation (non-exhaustive supervision) scenarios.
EviDent-CBCT inserts a discrete evidence record as a bottleneck layer between the image and the report text.
EviDent-CBCT uses an anatomy-aware network to map each CBCT scan into three levels of evidence: tooth level, global level, and tooth–inferior alveolar canal (tooth-IAC) relationship.
EviDent-CBCT's final report is generated by a deterministic renderer and a local language model that does not see the image, with the language model only having access to the evidence record and not the original imaging data.
For tooth-level evidence, EviDent-CBCT employs reliability-aware training, treating eligible unmentioned samples as down-weighted negative samples participating in training rather than simply treating them as negatives.
For unreported global evidence and tooth-IAC labels, EviDent-CBCT directly maintains an 'unknown' state rather than forcing an inference.
EviDent-CBCT designs metal-sensitive input channels that preserve intensity cues from dental materials such as fillings and implants.
Across three validation runs, EviDent-CBCT achieved a combined evidence set F1 of 0.666±0.006 and a RadFact-Lite-Dental logical F1 of 0.402±0.003.
The strongest controlled direct baseline achieved only 0.371±0.018 on the logical F1 metric, which is lower than EviDent-CBCT's 0.402.
In the ODIN 2026 challenge, EviDent-CBCT ranked second in automated evaluation and third in the blinded clinical arena comparison on the hidden test set.
1 more timeline events
All Facts (11)
As a newly released preprint, the results of EviDent-CBCT still await validation through larger-scale clinical trials.
50%UnverifiedEviDent-CBCT inserts a discrete evidence record as a bottleneck layer between the image and the report text.
50%UnverifiedEviDent-CBCT uses an anatomy-aware network to map each CBCT scan into three levels of evidence: tooth level, global level, and tooth–inferior alveolar canal (tooth-IAC) relationship.
50%UnverifiedEviDent-CBCT's final report is generated by a deterministic renderer and a local language model that does not see the image, with the language model only having access to the evidence record and not the original imaging data.
50%UnverifiedFor tooth-level evidence, EviDent-CBCT employs reliability-aware training, treating eligible unmentioned samples as down-weighted negative samples participating in training rather than simply treating them as negatives.
50%UnverifiedFor unreported global evidence and tooth-IAC labels, EviDent-CBCT directly maintains an 'unknown' state rather than forcing an inference.
50%UnverifiedEviDent-CBCT designs metal-sensitive input channels that preserve intensity cues from dental materials such as fillings and implants.
50%UnverifiedAcross three validation runs, EviDent-CBCT achieved a combined evidence set F1 of 0.666±0.006 and a RadFact-Lite-Dental logical F1 of 0.402±0.003.
50%UnverifiedThe strongest controlled direct baseline achieved only 0.371±0.018 on the logical F1 metric, which is lower than EviDent-CBCT's 0.402.
50%UnverifiedEviDent-CBCT is an Evidence-Bottlenecked framework for dental CBCT report generation, specifically designed for incomplete annotation (non-exhaustive supervision) scenarios.
50%UnverifiedIn the ODIN 2026 challenge, EviDent-CBCT ranked second in automated evaluation and third in the blinded clinical arena comparison on the hidden test set.
50%