Abstract:

Objective: To construct a deep learning model based on CT images for activity assessment of pulmonary tuberculosis lesions. Methods: A retrospective cohort of 102 cured pulmonary tuberculosis patients at Beijing Chest Hospital, Capital Medical University between December 2018 and December 2020 was included, CT data were collected before, during, and after treatment. Lesions were randomly divided into training and test sets with an 8∶2 ratio. Additionally, a prospective cohort of 72 cured pulmonary tuberculosis patients was enrolled between October 2021 and December 2022, CT datasets were collected for an independent validation set. A deep learning model was constructed through transfer learning using the Mask R-CNN architecture to achieve automatic lesion segmentation and activity determination. The model was trained based on three-dimensional lesion labels from the training set, and its performance in determining the activity of pulmonary tuberculosis lesions was evaluated in the test set and independent validation set by calculating the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Results: A retrospective cohort of 102 cured pulmonary tuberculosis patients who met the criteria was included, and a total of 770 CT imaging data were collected; 332 lesions were active, and 464 lesions were inactive. A prospective cohort of 72 cured patients with pulmonary tuberculosis was included, and a total of 540 CT imaging data were collected. The transfer learning-based Mask R-CNN deep learning model achieved an AUC of 87.5%, sensitivity of 85.7%, and specificity of 78.6% in the test set. In the independent validation set, the model obtained an AUC of 79.9%, sensitivity of 78.7%, and specificity of 75.0%. Conclusion: The transfer learning-based Mask R-CNN deep learning model has shown promising potential in predicting the activity of small-scale pulmonary tuberculosis lesions, could offer valuable scientific insights for rapid and automatic clinical decision-making.

Key words: Tuberculosis, pulmonary, Tomography, X-ray computed, Models, structural, Artificial intelligence