Abstract
Tuberculosis (TB) screening with chest radiography is attractive in resource-constrained settings, but high clean-test performance may conceal clinically important failure modes. We evaluated the backdoor robustness of a transfer-learned DenseNet-121 TB screening model using a Nigerian chest X-ray dataset from Aminu Kano University Teaching Hospital (AKTH), reformulated from four classes into TB-positive versus Non-TB. An external NIAID chest X-ray dataset was used for TB/normal pretraining before fine-tuning on the Nigerian dataset. We introduced a targeted bottom-right border trigger by poisoning 5%, 10%, and 20% of TB-positive training images and relabeling them as Non-TB. The clean model achieved 99.50% accuracy, 99.33% sensitivity, 99.56% specificity, and 0.9995 AUROC. Backdoor vulnerability increased with poisoning, reaching an ASR of 86.67% at 20% poisoning despite 98.17% clean accuracy. Repeated experiments across five random seeds confirmed the dose-dependent vulnerability, with 20% poisoning producing an ASR of 80.40 ± 7.78%. An 8% border-crop defense reduced ASR from 86.67% to 8.67% and improved triggered sensitivity from 13.33% to 91.33%, although clean sensitivity decreased. Grad-CAM demonstrated trigger-focused attention under attack and restored thoracic attention after cropping. These findings show that clean validation alone is insufficient for trustworthy TB screening AI and support lightweight robustness evaluation in resource-constrained settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_008.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_008_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=vJgrFWElR2
BibTex
@InProceedings{JonWis_Trustworthy_MICCAISAT2026,
author = { Jonathan, Wisdom AND Fidelis, Dabrilagha AND Collins, Emevor Edisemokumor},
title = { { Trustworthy TB Screening: Backdoor Vulnerability and Lightweight Defense in Chest X-Ray AI } },
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}
}
