Abstract
Automated CBCT-to-report generation is increasingly framed as a language-model problem and scored by overlap with reference reports. We ask whether a stronger report writer produces a better grounded report, or merely a better scoring one. On ODIN 2026 ToothFairy4 (627 CBCT volumes, no masks released) we fix the vision half, a public pretrained segmenter yielding deterministic coarse findings, and vary only the writer: (C) a deterministic template floor, (A) a frozen 32B instruction model, (B) the same model fine-tuned with QLoRA. Fine- tuning wins decisively: clinical-keyword-F1 0.359 →0.369 →0.500. A two-part faithfulness audit shows the gain is confabulation, not grounding: Mode B invents reference-style specifics the findings cannot support (only ∼47% of its tooth claims are correct) and overrides the base model’s faithfulness, echoing the grounded mandibular-canal signal in 29% of true cases (vs. 89% frozen) while fabricating canal risk in 56% of safe cases. Reproducing the challenge’s own RadFact-based composite, Mode B leads it too (0.29 vs. ≤0.10): entailment against the reference report is not grounding in the scan. We therefore introduce faithful-by-construction generation, deterministic schema gates over the frozen model with no training, which restores grounding (canal recall 1.00 at zero false alarms; composite 0.94–0.99 vs. 0.02 for QLoRA) at frozen-model fluency. Coarse findings, not the writer, are the information ceiling. We also report a caveat relevant to all participants: the maxilla is cropped out of the field of view in ∼70% of scans. Code is available at https://github.com/ VanchhayNheng/toothfairy4-probe.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ODIN_025.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=1IQzphMnIk
BibTex
@InProceedings{VanNhe_Fluency_MICCAISAT2026,
author = { Vanchhay, Nheng AND Khan, Ziaullah AND Mozumder, Md Ariful Islam AND Kim, Hee-Cheol},
title = { { Fluency Is Not Faithfulness: Diagnosing and Repairing Confabulation in LLM Report Generation on ToothFairy4 CBCT } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17270},
month = {pending},
page = {pending}
}