T-SPOT with CT image analysis based on deep learning for early differential diagnosis of nontuberculous mycobacteria pulmonary disease and pulmonary tuberculosis

Abstract

Objectives: This study aimed to establish a diagnostic algorithm combining T-SPOT with computed tomography image analysis based on deep learning (DL) for early differential diagnosis of nontuberculous mycobacteria pulmonary disease (NTM-PD) and pulmonary tuberculosis (PTB). Methods: A total of 1049 cases were enrolled, including 467 NTM-PD and 582 PTB cases. A total of 320 cases (160 NTM-PD and 160 PTB) were randomized as the testing set and were analyzed using T-SPOT combined with the DL model. The testing cases were first divided into T-SPOT-positive and -negative groups, and the DL model was then used to separate the cases into four subgroups further. Results: The precision was found to be 91.7% for the subgroup of T-SPOT-negative and DL classified as NTM-PD, and 89.8% for T-SPOT-positive and DL classified as PTB, which covered 66.9% of the total cases, compared with the accuracy rate of 80.3% of T-SPOT alone. In the other two remaining groups, where the T-SPOT prediction was inconsistent with the DL model, the accuracy was 73.0% and 52.2%, separately. Conclusion: Our study shows that the new diagnostic system combining T-SPOT with DL based computed tomography image analysis can greatly improve the classification precision of NTM-PD and PTB when the two methods of prediction are consistent.

Keywords

• Nontuberculous mycobacteria pulmonary disease
• Pulmonary tuberculosis
• Computed tomography
• T-SPOT
• Artificial intelligence
• Deep learning