Validation of an AI-Driven Sputum Analysis Platform for Accurate and Reliable Tuberculosis Detection in South-Western...

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 Ugandan regional referral laboratory.

A retrospective laboratory-based cross-sectional study used 512 archived digitized ZN-stained sputum smear images from Mbarara Regional Referral Hospital. The dataset contained 256 AFB-positive and 256 AFB-negative images and was partitioned by stratified sampling into training (n=358), validation (n=77), and an internal hold-out test set (n=77). ResNet-50 V2, DenseNet-121, and Swin Transformer-Tiny models were trained using ImageNet-pretrained weights in PyTorch. Performance was assessed using sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), F1-score, and ROC-AUC.

On the internal test set, Swin Transformer-Tiny achieved 97.40% accuracy, 97.37% sensitivity, 97.44% specificity, 97.37% PPV, 97.44% NPV, 97.37% F1-score, and ROC-AUC 0.9939. DenseNet-121 achieved 88.31% accuracy, 84.21% sensitivity, 92.31% specificity, and ROC-AUC 0.9730. ResNet-50 V2 achieved 85.71% accuracy, 81.58% sensitivity, 89.74% specificity, and ROC-AUC 0.9555. Wilson 95% confidence intervals were calculated for the threshold-based metrics.

The models demonstrated promising internal performance for classification of microscopy-defined AFB status. Swin Transformer-Tiny had the highest observed performance, but the small, balanced, single-centre internal test set and use of routine microscopy labels rather than an independent microbiological reference standard limit clinical interpretation. The platform should therefore be considered an AI-assisted microscopy reading aid rather than an autonomous tuberculosis diagnostic system. Prospective, multicentre evaluation using patient/specimen-level separation, independent microbiological reference standards, and multiple blinded readers is required before clinical deployment.