Abstract:
Mycobacterium Tuberculosis (M. TB) bacteria causes Tuberculosis, which leads a human to death. The World Health Organization declared TB a highly infectious disease and reported that 1.7 million individuals die yearly. WHO recently constructed a target to reduce TB by 80% by 2030 and 90% by 2035 and passed a recommendation to employ an artificial intelligence-based computer-aided diagnosis process. It is challenging for the radiologist to differentiate TB and Lung Cancer from CXR imaging, as both imitate each other. The screening process is relatively difficult to detect early from the most widely-used "Chest X-ray" imaging of individuals that requires a significant degree of accuracy. This study aimed to train a convolution neural network model from scratch to detect tuberculosis from chest X-ray images and compare its performance with the transfer learning technique using mobile net pre-trained and multiple-layer CNN models. This study has used the publicly available datasets from Kaggle of chest radiographs, trained a CNN model from scratch, and utilized a multi-layer CNN model and mobile net to classify TB and normal cases. Our proposed model is mobile net transfer learning. The image-preprocessing, data augmentation and optimization has been done by using rmsprop in the proposed model. The performance of the model was evaluated using accuracy. The result showed that all the proposed models had accepted accuracy for two-class classifications, with our proposed CNN architecture achieving 96.63% and our lowest accuracy being 90.00%, which was obtained from the multilayer CNN model.
