Multitask Deep Convolutional Neural Network with Attention for Pulmonary Tuberculosis Detection and Weak Localization of Pathological Manifestations in Chest X-Ray

Pulmonary tuberculosis (PTB) is a highly fatal bacterial infection that affects the lungs. Chest radiography is a commonly used technique for PTB diagnosis. Interpreting chest X-ray images for features like cavitation, consolidation, and nodules poses challenges due to low contrast between lesions and surrounding tissue, and the complexity of identifying features for intricate disorders. To address these challenges, researchers have proposed using deep learning techniques to detect and mark areas of TB infection in chest X-rays. However, fully supervised semantic segmentation requires massive large pixel-by-pixel labeled images, which is time-consuming, expensive, and subjective. As a result, there is growing interest in weak localization techniques, a method identifying disease pathologies without pixel-level labeling. Hence, this study focuses on developing a deep learning model for weakly supervised segmentation and localization of radiographic manifestations of PTB from chest X-rays (CXR), using commonly used public datasets for TB identification. We proposed multi-scale attention using the DenseNet-121 model as a backbone network. First, a class activation map is calculated at different levels of the backbone network using the last feature map and the global average pooling at each specific level. Finally, the class activation map is combined using a convex combination and passed to the sigmoid functions. This approach is powerful for classifying and localizing disease pathology in CXR. We achieved a localization accuracy of 83% for T (IoU) = 0.1 and the classification AUC, accuracy and \(F_1\) score are 98%, 98%, and 97% respectively. This result indicates the model has a promising performance in both the classification and localization of PTB manifestations.