Document Type : Original Article
Authors
1 Dept. of Biostatistics, sch. of Public Health, Iran University of Medical Sciences, Tehran, Iran.
2 MA, MS, PhD, Assistant Professor, Department of Biostatistics, School of Public Health, & Minimally Invasive Surgery Research Center, Iran University of Medical Sciences, Tehran, Iran.
Abstract
Background: One of the best ways to reduce the spread of Tuberculosis (TB) is diagnose the disease using chest X-ray (CXR) images as a low-cost and affordable method. It faces two problems: the lack of adequate radiologist, and the possibility of misdiagnosis. That is why it is necessary to use accessible and accurate diagnosis system. This research seeks to design an accurate and accessible automatic diagnosis system based on deep learning that can solve diagnosis problems.
Methods: Six Convolutional Neural Networks InceptionV3, ResNet50, DenseNet201, MnasNet, MobileNetV3, and EfficientNet-B4 trained using Transfer learning, Adam optimizer20 training epochs, and also using the new, large, and accurate dataset TBX11K. The network could classify each image into one of the three groups of patients with TB, patients with lung disease other than TB, so the lung abnormality in their images is not caused by TB, and healthy people without TB and lung image does not show pulmonary abnormality.
Results: In the testing step, the networks achieved very high performance. The EfficientNet-B4 network outperformed the other networks with sensitivity %97.1, specificity %99.9, and accuracy %99.5. It has also a higher performance than previous studies in the field of TB diagnosis using CXR images by Convolutional Neural Networks.
Conclusion: This research showed with having access to high quality, large datasets, and standard training, it is possible to entrust the diagnosis of TB with high confidence using medical images to computers and artificial neural networks, which in this research were able to achieve accuracies higher than 99%.
Keywords
Main Subjects
)