Abstract
Report generation for 3D dental cone-beam computed to- mography (CBCT) remains a largely manual process. Automating this process is difficult because of the heavy computational burden and ex- cessive token consumption required to process raw 3D volumetric data, as well as the specialized nature of dental CBCT that general-purpose models lack training to interpret for fine-grained pathological features. To address this, we propose an anatomy-grounded, schema-constrained pipeline that uses segmentation as a spatial index to convert the 3D vol- ume into informative multi-view visual inputs, including 3D renderings, panoramic reconstructions, and multi-slice composites for detailed eval- uation. These visual inputs, combined with a clinically defined schema, are fed into a LoRA fine-tuned vision-language model (VLM) for vi- sual question answering (VQA). The structured outputs undergo rule- based post-processing and are then mapped to clinical phrasing snippets to assemble the final report. On the ODIN 2026 / ToothFairy4 bench- mark, our fine-tuned pipeline reaches a local held-out clinical (RadFact F1) score of 0.444 and a final score of 0.409, and an official RadFact F1 of 0.392 and final score of 0.361 on the hidden test set, and a per- finding breakdown shows the binding constraint is detecting fine-grained, visually ambiguous findings. The source code is publicly available at https://github.com/CC-AIDent-Vienna/v2v2R.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ODIN_challenges_024.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/ODIN_challenges_024_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=PcZnThoeWm
BibTex
@InProceedings{SinShi_From_MICCAISAT2026,
author = { Singh, Shivam AND Lu, Chen AND Sagl, Benedikt},
title = { { From Voxels to Views to Reports: A Segmentation-Guided VLM Pipeline for CBCT Report Generation } },
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}
}