OrthoScribe: Assets for Training-Free Multimodal Orthodontic Report Generation
Year: 2026
Type: software
Abstract: Released assets for OrthoScribe, a training-free multimodal pipeline for orthodontic report generation from intraoral photographs and 3D intraoral scans, submitted to the MICCAI 2026 ODIN Bite2Text challenge (Task 2). This archive (orthoscribe_assets.zip) contains the components required to build and run the inference container:- report_bank_full.json: 781 training-set orthodontic reports used as the retrieval bank- train_img_embeds_full.npz: DINOv2 image embeddings for the 781 training cases- train_geom_full.json: per-case overjet and overbite measurements for the training set- geom_stats.json: z-score normalisation constants for geometry-based re-ranking- baselines.json: constant medoid report used as a fallback- dinov2/: the self-supervised DINOv2 image encoder (config, preprocessor, safetensors weights) Code: https://github.com/IUCompPath/MICCAI_Challenge_Bite2TextThe ODIN 2026 Bite2Text dataset is released by the challenge organisers and is not included here. MD5 (orthoscribe_assets.zip): AD5536FA0928157B6DDC3BBEF7A3119C
Authors: Kachole, Sanket, Bakas, Spyridon
Institutions: Indiana University Health, Indiana University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204272478