Abstract:

Objective To evaluate the value of feature pyramid network (FPN) for screening pulmonary tuberculosis on the thoracic radiography images (hereinafter referred to as “chest images”).Methods In this retros-pective study, chest images of 490 pulmonary tuberculosis patients who were hospitalized in the Tuberculosis Hospital of Shaanxi Province from January 2016 to December 2017 were collected, as well as chest images of 100 healthy outpatient controls. In addition, chest images of 332 and 58 pulmonary tuberculosis patients separately from Shenzhen, China and Montgomery, Maryland, USA in the NIH public dataset were also included. FPN was employed to classify and localize the radiographs and lesions. Two radiologists from tuberculosis hospitals examined and labeled the chest images of pulmonary tuberculosis patients in the above data. After data alignment and augmentation, the annotated radiographs were sent to the FPN, and the FPN was trained to obtain the final detection model. Then, the performance and generalization ability of FPN were tested with independent dataset, and the sensitivity, specificity and accuracy of FPN in distinguishing chest images of pulmonary tuberculosis patients from healthy controls were analyzed according to the criteria of positive sputum smear and evaluation by experienced radiologists. Meanwhile, the sensitivity and false positive rate of FPN in localizing tuberculosis lesions were evaluated based on manually labeled lesions. For the lesion detection and localization in the images, the free response receiver operating characteristic (FROC) score was used to evaluate the performance of FPN.Results The sensitivity, specificity and accuracy of FPN in diagnosing pulmonary tuberculosis on the test sets were 96.0% (96/100), 76.0% (76/100) and 86.0% (172/200), respectively. A total of 226 lesions were labeled on 100 positive chest images, while 242 lesions were detected in FPN, with sensitivity and false positive rate of 87.6% (198/226) and 14.0% (34/242), respectively. The FROC localization score could be up to 88.0%. Conclusion FPN can effectively differentiate chest images of pulmonary tuberculosis patients from healthy controls and implement the localization of lesions, which provide a reference for the deep learning-based classification and lesion localization of tuberculosis.

Key words: Tuberculosis, pulmonary, Radiography thoracic, Chest X-ray, Artificial intelligence, Diagnosis, differential, Deep learning, Automatic screening