OBJECTIVES: To train and evaluate the performance of a detector for pulmonary tuberculosis and pulmonary cavity, using the Faster Region Convolutional Neural Network model. STUDY DESIGN: A cross-sectional study design was employed to describe the sensitivity, specificity, and accuracy of the Faster Region Convolutional Neural Network model for the detection of pulmonary tuberculosis and pulmonary cavity. SUBJECTS: Radiographs for the training dataset and testing dataset were acquired from the Picture Archiving and Communication System of the a general public hospital in Quezon City. SETTING: The setting of the study is a general public hospital in Quezon City, Philippines. OUTCOMES: The detector for pulmonary tuberculosis and pulmonary cavity was trained with the training dataset using the TensorFlow machine learning library, with the Faster-RCNN-lnception-V2 used as the base model. Detector findings on the testing dataset were compared and analyzed against findings of three board-certified radiologists. RESULTS: The detector achieved 92.11 % sensitivity, 87.1 % specificity, and 89% accuracy as a screening tool, and 84.04% sensitivity, 98.04% specificity, and 95.87% accuracy, as a locator of pulmonary tuberculosis and cavity. CONCLUSION This study is the first of its kind to demonstrate the feasibility of training a detector for pulmonary tuberculosis and pulmonary cavities using the Region Convolutional Neural Network model. Limitations and improvements to the detector may be addressed in future research.
Tuberculosis, Pulmonary ; Sensitivity and Specificity ; Neural Networks (Computer) ; Software