Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2125
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-21T22:02:37.165119Z
status: published
lens: g_space
sector: health
headline: Validation of an AI-Driven Sputum Analysis Platform for Accurate and Reliable Tuberculosis Detection in South-Western Uganda
dek: Ziehl-Neelsen (ZN) sputum smear microscopy remains an accessible method for detecting acid-fast bacilli (AFB), but its interpretation is dependent on specimen quality and trained microscopy personnel. Artificial intelligence (AI) may provide a microscopy reading aid that supports consistent identification of AFB-positive and AFB-negative smear images. To evaluate transfer-learning deep learning models for classification of AFB-positive and AFB-negative digitized ZN-stained sputum smear images generated in a Ugan…
gain_title: A Swin Transformer-Tiny model separated AFB-positive from AFB-negative Ziehl-Neelsen smear images with 97.40% accuracy on an internal test set, suggesting a possible microscopy reading aid.
problem_title: (none)
trace_subject: (none)
gain_reading: A Swin Transformer-Tiny model separated AFB-positive from AFB-negative Ziehl-Neelsen smear images with 97.40% accuracy on an internal test set, suggesting a possible microscopy reading aid.
gain_evidence: Swin Transformer-Tiny achieved 97.40% accuracy, 97.37% sensitivity, 97.44% specificity
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers trained three pretrained deep learning models on 512 archived digitized Ziehl-Neelsen sputum smear images from Mbarara Regional Referral Hospital in south-western Uganda. The images were split evenly between AFB-positive and AFB-negative, and the models were tested on a held-out set of 77 images.

The result points to a possible way to support consistent smear reading where trained microscopy staff and specimen quality limit interpretation. Yet the evidence is early: the test set was small and from one centre, and the labels came from routine microscopy, so the system is not shown to be a stand-alone diagnostic tool.
limitation: The small, balanced, single-centre internal test set and reliance on routine microscopy labels rather than an independent microbiological reference standard limit clinical interpretation.
tag: Evidence-backed gain
key_points: The dataset held 256 AFB-positive and 256 AFB-negative images, split into training (358), validation (77) and test (77) sets. | Swin Transformer-Tiny reached ROC-AUC 0.9939 on the internal test set. | DenseNet-121 reached 88.31% accuracy and ResNet-50 V2 reached 85.71% accuracy. | The authors describe the platform as an AI-assisted reading aid, not an autonomous diagnostic.
rundown: The study was a retrospective, laboratory-based cross-sectional analysis. ResNet-50 V2, DenseNet-121 and Swin Transformer-Tiny were trained with ImageNet-pretrained weights in PyTorch, and performance was judged by sensitivity, specificity, accuracy, PPV, NPV, F1-score and ROC-AUC, with Wilson 95% confidence intervals for threshold-based metrics.

Swin Transformer-Tiny posted sensitivity of 97.37% and specificity of 97.44%, while DenseNet-121 had sensitivity of 84.21% and specificity of 92.31%. ResNet-50 V2 had sensitivity of 81.58% and specificity of 89.74%, so the transformer model led on every reported measure.

The authors call for prospective, multicentre evaluation with patient or specimen-level separation, independent microbiological reference standards and multiple blinded readers before any clinical deployment.
sources:
- peer_reviewed | Medical Devices: Evidence and Research | https://doi.org/10.2147/mder.s639007 | 2026-10-15
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

55ae425adbeb91ee25dd1a14c880b578599555f8b5be5f5901ba0c866b390d71

previous

0000000000000000000000000000000000000000000000000000000000000000