TB detection on chest radiographs: The Mn AI chest X-ray model achieves a higher AUC than Google's deep learning system on the Shenzhen and Montgomery tuberculosis benchmarks

Chest radiography (CXR) is a first-line tool for tuberculosis (TB) screening and triage, but radiologists are scarce in many of the settings where TB is most common. Deep learning systems (DLS) can help by flagging abnormal radiographs for prompt review. Here we evaluate the Mn Al chest X-ray model, which classifies radiographs as normal or abnormal, on the two most widely used public TB benchmarks, the Shenzhen (n = 662; 336 TB) and Montgomery (n = 138; 58 TB) collections released by the US National Library of Medicine. Neither collection was used for training or tuning, a duplicate search found no overlap with training images, and the decision threshold was fixed in advance on a separate set of 600 radiographs. The model achieved an area under the receiver operating characteristic curve (AUC) of 0.998 (95% CI 0.994-1.000) on Montgomery and 0.965 (0.950-0.977) on Shenzhen. At the pre-specified threshold, sensitivity and specificity were 93.1% and 100% on Montgomery and 83.9% and 96.9% on Shenzhen. For reference, Google's published DLS for normal-versus-abnormal CXR classification reported AUCs of 0.97 (0.94-0.99) and 0.95 (0.93-0.97) on posteroanterior subsets of the same two collections. On Montgomery, the Mn Al model achieved the highest AUC of the published systems compared, ahead of both Google's DLS (with non-overlapping 95% confidence intervals) and a TB-specific system trained on 6,768 TB radiographs (0.996); on Shenzhen it also exceeded Google's result (0.965 against 0.95), while running on a single consumer graphics card or, more slowly, on a standard desktop CPU. These results position the model as a strong candidate for TB triage in screening programmes.