Abstract
MMDental Task 3 predicts seven clinical-report fields from 3D dental CBCT with only 50 labeled reports. We formulate this small- data problem as constrained field retrieval rather than free-text gen- eration, so each output field is traceable to one labeled training re- port. The proposed framework combines a MedicalNet 3D ResNet-34 with a SwinUNETR-derived encoder. Image-only contrastive adaptation uses all 250 official training CBCTs, including the unlabeled subset, before supervised learning maps each encoder representation to seven field-specific slices of a 168-D latent report vector. At inference, each branch independently ranks the fixed field-specific candidate banks us- ing visual and predicted-report similarity. A medication-preserving gate fixes Diagnosis and Handle to the MedicalNet branch and admits a Swin candidate for another field only when the two candidates contain the same recognized medication-pattern set. Official online evaluation yields a Weighted_Score of 0.235259, diagnosis-code F1 of 0.161000, treatment-action F1 of 0.346893, medication F1 of 0.470000, and com- plete nonempty output for all evaluated cases. The results support au- ditable retrieval as a practical alternative to unconstrained generation in a highly limited supervision regime. The gate preserves recognized lexical mentions but does not certify clinical safety.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ODIN_challenges_019.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=et8vccbqRE
BibTex
@InProceedings{WanLai_MedicationSafe_MICCAISAT2026,
author = { Wang, Lai AND Yue, Changpeng AND Bao, Wentao AND Shao, Junhao},
title = { { Medication-Safe Dual-Encoder Retrieval for Structured Dental Report Prediction from 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}
}