A Generalized Deep Learning Approach for Evaluating Secondary Pulmonary Tuberculosis on Chest Computed Tomography

Abstract

Background Pulmonary tuberculosis (PTB) is a global health problem and remains the leading infectious cause of death worldwide. Differentiation between secondary PTB and non-tuberculous (non-TB) pneumonia is important for patient isolation and treatment, but can be difficult to determine clinically and radiologically. We proposed the application of deep learning to chest computerized tomography (CT) to assist doctors in detecting and differentiating PTB from non-TB pneumonia in an expedient, non-invasive, and reproducible manner. Methods: We retrospectively collected a dataset containing 1,124 CT scans from 923 PTB and non-TB pneumonia based on their pathological reports of lung biopsy and clinical information and 201 patients without pulmonary infiltrate from West China Hospital between 2012 and 2018. Randomly selected parts of this dataset (WCPR dataset) were used to develop, train, internally validate and test the algorithm. Patients in the WCPR dataset (PTB, n=439; non-TB pneumonia, n=484; normal, n= 201) were randomly assigned in three non-overlapping sets: training set, n=866; validation set, n=108; and test set, n=150. An additional dataset from NIH TB Portal31 comprising of cases from three countries (Belarus, n = 274; Romania, n = 43; Moldova, n = 10) was used to validate externally the algorithm's ability to identify PTB. A convolutional neural network of Inception-Res-Net-v230 was trained and tested on the entire chest CT to mimic real life application. The performance of our algorithm was compared to three trained radiology/pulmonology physicians. Findings For differentiating pulmonary infiltrates, the algorithm achieved 99·3% accuracy (149 out of 150), 99.0% sensitivity, and 100·0% specificity. For identifying PTB, the algorithm achieved 82·0% accuracy (123 out of 150), 95·9% sensitivity, and 75·2% specificity. For identifying non-TB pneumonia, the algorithm achieved 81·3% accuracy (122 out of 150), 52·7% sensitivity, and 97·9% specificity. This mostly outperformed our human readers for PTB identification, who averaged up to 81·1% accuracy, 70·8% sensitivity, and 86·1% specificity. Our algorithm identified 287 out of 327 PTB (87·8% accuracy) cases in NIH TB Portal Dataset from other countries. Interpretation: Our deep-learning-based algorithm successfully differentiated abnormal from normal chest CTs, as well as PTB from non-TB pneumonia cases and thus allows real world applicability. Early identification of PTB from non-TB pneumonia can help control outbreaks through isolation and early appropriate treatment. The application of our algorithm could expedite the identification of PTB, thereby improving clinical outcomes. Our datasets and algorithm used in this study will be publicly available to facilitate world-wide adoption. Funding Statement: The authors declare: None. Declaration of Interests: The authors declare: None. Ethics Approval Statement: This study was approved by the Institutional Review Board of West China Hospital (approval No. 2019-148) and Icahn School of Medicine at Mount Sinai (approval No. GCO#1: 19-0569(0001) ISMMS), and the patients’ written consents were waived.