Tuberculosis (TB) is a major health threat in the developing countries. Many patients die every year due to lack of treatment and error in diagnosis. Developing a computer-Aided diagnosis (CAD) system for TB detection can help in early diagnosis and containing the disease. Most of the current CAD systems use handcrafted features, however, lately there is a shift towards deep-learning-based automatic feature extractors. In this paper, we present a potential method for tuberculosis detection using deep-learning which classifies CXR images into two categories, that is, normal and abnormal. We have used CNN architecture with 7 convolutional layers and 3 fully connected layers. The performance of three different optimizers has been compared. Out of these, Adam optimizer with an overall accuracy of 94.73 and validation accuracy of 82.09 performed best amongst them. All the results are obtained using Montgomery and Shenzhen datasets which are available in public domain. © 2017 IEEE.