Abstract

We address report generation from oral imaging for both tasks of the ODIN 2026 challenge: maxillofacial reporting from a cone- beam CT (CBCT) volume of the jaws (ToothFairy4, Task 1) and or- thodontic reporting from a 3D intraoral surface scan together with 2D intraoral photographs (Bite2Text, Task 2). Both inputs pair 3D geome- try with 2D images. A fixed 7-billion-parameter vision-language model is trained on natural 2D images and videos and cannot ingest a CBCT volume or a triangle mesh directly, so the input representation is the primary lever. We render the 3D data into the 2D reconstructions clin- icians use for these findings and additionally pass the volume itself as video frames, so the model receives both the diagnostic views and the un- derlying acquisition. We present RGBite, which pairs that representation with a structured output: rather than generating the report, the model predicts a dense 32-slot dental chart from which the report is rendered deterministically. Both are instantiated on attenuation for Task 1 and surface geometry for Task 2, with a Qwen2.5-VL-7B backbone LoRA- fine-tuned identically for both. On a held-out split the richer representa- tion is worth +0.031 FinalScore at fixed supervision, while rebalancing the chart’s non-tooth lines by inverse frequency improves those fields but costs an equal amount of tooth accuracy and −0.030 on the report. We report official challenge scores for both tasks alongside the local eval- uation, and a judge-free per-finding breakdown locating the remaining error in restorations and pathology rather than dentition.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ODIN_challenges_023.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=r4wQfRfDx9

BibTex

@InProceedings{SpiSör_RGBite_MICCAISAT2026,
        author = { Spiegel, Sören AND Bellmann, Louis AND Nielsen, Maximilian},
        title = { { RGBite: Vision-Language Report Generation for the ODIN 2026 Challenge } },
        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}
}


back to top