Abstract
We describe our submission to the tuberculosis (TB) Diagnosis task (Task 2) of the TREAT-MMTB-2026 challenge, which targets binary TB vs. Normal classification of chest X-ray (CXR), evaluated by the binary F1 score with TB as the positive class. Our encoder is a ViT-L/16 backbone adapted to CXR by vision–language pretraining (VLP) on MIMIC-CXR at 512 × 512, the challenge input resolution. Its patch tokens are attention-pooled and fed to a small MLP head. The released data are almost perfectly separable—an F1 score of 1.0 on a challenge-internal split and 0.9978 on the released test split—so no released measurement could rank our models. Every design choice targets generalization to distribution shift instead of internal fit. We widen the training set with publicly available external CXR datasets, diversify the ensemble in training-data composition and augmentation policy, and select the decision threshold on a validation set drawn from external sources. The submitted model is an ensemble of four classifiers, trained on five datasets or on those plus the Shenzhen and Montgomery tuberculosis sets, each configuration with and without image augmentation—flip-averaged at inference and thresholded at 0.35. It reaches an F1 score of 0.8642 on the organizers’ external evaluation. The gap between the released splits and that evaluation is the central empirical result of this work, and the only signal that could distinguish our configurations. Our model placed second in Task 2.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TREAT_MMTB_005.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=CZZCTepmMj
BibTex
@InProceedings{LeeSeo_Generalizable_MICCAISAT2026,
author = { Lee, Seongeun AND Yun, Hannah AND Jeong, Taejin AND Jeon, Mingyeong AND Park, Junhyun AND Kim, Hyunwoong AND Park, Jonggwon},
title = { { Generalizable Tuberculosis Classification on Chest X-rays through Multi-Source Curation and Model Ensembling } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}