Distinguishing infectivity in patients with pulmonary tuberculosis using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• Pulmonary tuberculosis (PTB) remains a significant global health challenge.
• Accurate assessment of PTB infectivity is crucial for effective disease control.
• Current methods for assessing PTB infectivity can be time-consuming and resource-intensive.

Purpose of the Study:

• To develop and evaluate a deep learning model, TBINet, for distinguishing infectivity in PTB patients using computed tomography (CT) images.
• To compare the performance of TBINet against conventional deep learning models.
• To enhance the interpretability of the deep learning model's predictions.

Main Methods:

• A dataset of 925 PTB patients from four centers was curated and labeled for weak and strong infectivity based on sputum smear results.
• The TBINet model was trained on this dataset and its performance was compared to the 3D ResNet model.
• Gradient-weighted class activation mapping (Grad-CAM) was employed for model explainability to identify lesion activation sites in CT images.

Main Results:

• TBINet achieved superior performance, with an area under the curve (AUC) of 0.819 on the validation set and 0.753 on the external test set.
• Grad-CAM analysis indicated that CT features such as consolidation, voids, upper lobe involvement, and enlarged lymph nodes correlated with higher PTB infectivity.
• The model demonstrated better accuracy in identifying infectious patients compared to existing deep learning methods.

Conclusions:

• The study demonstrates the feasibility of utilizing CT images and deep learning for assessing PTB infectivity.
• TBINet offers a promising, non-invasive tool for identifying infectious PTB patients.
• This approach could potentially aid in early detection and management strategies for PTB.