Deep learning-based differentiation of non-tuberculous mycobacterial lung disease and pulmonary tuberculosis using chest CT

Background Differentiating non-tuberculous mycobacterial lung disease (NTM-LD) from pulmonary tuberculosis (PTB) remains a significant clinical challenge owing to their overlapping clinical and imaging features, despite markedly different therapeutic strategies. Methods A total of 409 patients with microbiologically confirmed diagnoses were retrospectively enrolled and randomly divided into a training set ( n = 329; NTM-LD: 171, PTB: 158) and an independent test set ( n = 80; NTM-LD: 41, PTB: 39). After lung segmentation with nnU-Net, images were intensity-normalized and resampled to 256 × 256 × 128 voxels. A 3D ResNeXt-based classifier was developed and compared against six mainstream deep learning architectures: ResNet, SENet, DenseNet, ShuffleNet, Transformer, and Swin Transformer. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. Results The proposed 3D ResNeXt model achieved the highest performance, with an AUC of 0.89 and accuracy of 0.89 on the training set, and an AUC of 0.83 and accuracy of 0.84 on the independent test set. DeLong’s test confirmed statistically significant superiority over all six comparator architectures on the test set (all p < 0.05). Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations highlighted disease-specific features, including nodular bronchiectasis and tree-in-bud opacities in NTM-LD, as well as thick-walled cavitary lesions in PTB. Conclusion The 3D ResNeXt model demonstrated superior and interpretable performance in differentiating NTM-LD from PTB on chest CT. It holds promise as a valuable clinical decision-support tool, although prospective multicenter validation is warranted.