Abstract:

Objective: To develop and assess the effectiveness of various deep learning networks based on CT imaging for distinguishing nontuberculous mycobacterial lung disease (NTM-LD) from pulmonary tuberculosis (PTB). Methods: A retrospective analysis was performed on chest CT images from patients with PTB and NTM-LD at the Tuberculosis Department of Shanghai Pulmonary Hospital between October 2019 and December 2022. The data were divided into two groups: the PTB group (197 cases) and the NTM-LD group (212 cases). Each group was split into training and testing sets in an 8∶2 ratio. Diagnostic models were developed using several deep learning networks, including ResNeXt, ResNet, Transformer, SENet, ShuffleNet, Swin-Transformer, and DenseNet. The diagnostic performance of the models was assessed based on the area under the curve (AUC), accuracy, sensitivity, and specificity. The performance of the optimal model was then compared with that of three radiologists with varying years of experience, using the testing set. Results: The AUC values of the ResNeXt, ResNet, Transformer, SENet, ShuffleNet, Swin-Transformer, and DenseNet models on the training set were 0.894, 0.839, 0.864, 0.816, 0.841, 0.831, and 0.829, respectively. On the test set, the AUC values were 0.826, 0.816, 0.817, 0.811, 0.784, 0.771, and 0.735, respectively. In the test set, the ResNeXt model, with an AUC of 0.826, outperformed both a moderately experienced doctor (AUC: 0.679) and a less experienced doctor (AUC:0.663), with Z values of 2.035 and 2.242, respectively (P<0.05). Conclusion: The deep learning network model based on chest CT images is a rapid, simple, and non-invasive diagnostic tool, demonstrating excellent performance in distinguishing NTM-LD from PTB.

Key words: Mycobacterium infections, Artificial intelligence, Tomography, X-ray computed, Diagnosis,differential