Tuberculosis detection with customized CNN and oversampling techniques: a deep learning approach

Rapid, sensitive and accurate tests are critical for contemporary health care. Tuberculosis has been with us since time immemorial, and although being treatable and curable, it remains one of the world’s most infectious malady, killing millions each year. Significant advances have been made in tuberculosis diagnostic technologies however, they are not adequately portable, rendering them inapplicable in remote, rural areas where they are most required. They are costly, lack the essential sensitivity or precision and are frequently susceptible to error. In this regard, an automated system that can analyse and diagnose pulmonary TB by using Chest X-Rays (CXR) can prove to be a cost effective, dependable and quick alternative to manual diagnostic methods. In our work, we propose a customized CNN architecture having three convolution layers and three max pooling layers for accurately classifying the CXRs into TB-infected and normal classes. To deal with the issue of unbalanced classes in the TB CXR dataset, we use different oversampling techniques such as weighted averaging, SMOTE, ADASYN and Borderline SMOTE. To ensure a comprehensive assessment, we extended our research to further employ two widely recognized pre-trained CNNs; DenseNet169 and ResNet50 to gauge their performance on the same task. Grad-CAM has been applied to interpret the decision-making process of the models and correctly identify the regions of interest. Various evaluation metrics employed show very encouraging results. The proposed customized CNN model along with ADASYN outperforms DenseNet and ResNet models in terms of computation time as well.