Diagnosing Tuberculosis is crucial for proper treatment since it is one of the top 10 causes of deaths worldwide. Considering a computer-aided approach based on intelligent pattern recognition on chest X-ray with Convo- lutional Neural Networks, this work presents the proposition, training and test results of 9 different architectures to address this task as well as two ensem- bles. The highest performance verified reaches accuracy of 88.76%, surpassing human experts on similar data as previously reported by literature. The experi- mental data used comes from public medical datasets and comprise real-world examples from patients with different ages and physical characteristics, what favours reproducibility and application in practical scenarios.