Abstract
Tuberculosis remains the deadliest single infectious disease worldwide, with automated chest X-ray (CXR) screening being one of the few scalable routes to diagnosis where radiologists are scarce. A persistent obstacle is shortcut learning classifiers that appear near-perfect on held-out data from the same acquisition source but collapse on data from a different institution. We present a TB/Normal classifier built on two frozen self-supervised CXR foundation models, RAD-DINO and CheXFound, adapted with low-rank adaptation (LoRA) rather than full fine-tuning. Because a genuine acquisition-domain signal is present in the training data, we validate with leave-one-imaging-modality-out (LOMO) cross-validation rather than a stratified split, which would silently reward exploiting that signal. Lung-field cropping, multi-level image augmentation, random convolutions, and Fourier amplitude perturbation are applied throughout. Among five trained backbones, a probability-averaged ensemble of the two foundation models achieves the best internal validation F1 (0.9929), exceeding the full five-backbone ensemble (0.9890). We further evaluate on the public Shenzhen and Montgomery cohorts, where F1 falls to 0.7707 and 0.8348, respectively. This gap indicates that same-source validation — even under the comparatively strict LOMO protocol — substantially overestimates cross-institution performance, and we argue that closing this gap, rather than further internal-metric improvement, is the central open problem for deployable TB screening.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TREAT_MMTB_011.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=uRjWTUT6Mi
BibTex
@InProceedings{HwaEun_DomainAware_MICCAISAT2026,
author = { Hwang, Eunchan},
title = { { Domain-Aware Tuberculosis Screening with a LoRA-Adapted Chest X-Ray Foundation Model } },
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}
}