| Abstrakt: |
Tuberculosis is a dangerous disease to human life, and we need a lot of attempts to stop and reverse it. Significantly, in theCOVID-19 pandemic, access to medical services for tuberculosis has become very difficult. The late detection of tuberculosis could lead to danger to patient health, even death. Vietnamis one of the countries heavily affected by the COVID-19 pandemic, andmany residential areas as well as hospitals have to be isolated for a long time. Reality demands a fast and effective tuberculosis diagnosis solution to deal with the difficulty of accessingmedical services, such as an automatic tuberculosis diagnosis system. In our study, aiming to build that system, we were interested in the tuberculosis diagnosis problem from the chest X-ray images of Vietnamese patients. The chest X-ray image is an important data type to diagnose tuberculosis, and it has also received a lot of attention from deep learning researchers. This paper proposed a novel method for tuberculosis diagnosis and visualization using the deeplearning approach with a large Vietnamese X-ray image dataset. In detail, we designed our custom convolutional neural network for theX-ray image classification task and then analyzed the predicted result to provide visualization as a heat-map. To prove the performance of our network model, we conducted several experiments to compare it to another study and also to evaluate it with the dataset of this research. To support the implementation, we built a specific annotation system for tuberculosis under the requirements of radiologists in the Vietnam National Lung Hospital. A large experiment dataset was also from this hospital, and most of this data was for training the convolutional neural network model. The experiment results were evaluated regarding sensitivity, specificity, and accuracy. We achieved high scores with a training accuracy score of 0.99, and the testing specificity and sensitivity scores were over 0.9. Based on the X-ray image classification result, we visualize prediction results as heat-maps and also analyze them in comparison with annotated symptoms of radiologists. [ABSTRACT FROM AUTHOR] |
