Abstract
Chest radiograph AI for tuberculosis (TB) is usually judged on clean digital images, yet many resource-constrained screening pathways depend on scanned films, light-box photographs, compressed transfers, and inconsistent exposure. We stress-tested three TBX11K classifiers (ResNet50, DenseNet121, EfficientNet-B0) with deterministic paired corruptions that mimic ten digitisation artifacts at three severities. Model thresholds were calibrated once on clean validation images to preserve sensitivity, then frozen. Clean performance was nearly saturated: ROC AUC 0.9994–0.9998 and AUPRC 0.9959–0.9981. Under corruptions, however, the same systems separated into different operational risks. EfficientNet-B0 was dominated by false positive overload; severe Gaussian noise reduced AUPRC from 0.9959 to 0.0892 and added 1294 false positives, while severe low contrast added 1456 false positives. ResNet50 was more conservative but missed TB under severe defocus blur and under-exposure, adding 136 and 106 false negatives. DenseNet121 was most balanced on average, though severe under-exposure still added 101 false negatives. Grad-CAM++ stability analysis showed matching saliency drift, especially for EfficientNet-B0 under noise and ResNet50 under blur. We argue that TB CAD papers for low-resource deployment should report paired artifact tests, calibrated thresholds, absolute FP/FN shifts, and explanation stability, not clean AUC alone.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_010.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_010_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=t4Hsk7TjmT
BibTex
@InProceedings{SidFre_Clean_MICCAISAT2026,
author = { Sidume, Freedmore AND Muchuchuti, Stewart AND Chengetanai, Gibson AND Tamukate, Rabson AND Ntwaetsile, Kushatha},
title = { { Clean AUC Is Not Enough: Architecture-Specific Failure Modes of Tuberculosis Classifiers under Digitisation Artifacts } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}
