Tuberculosis (TB) is among the most common communicable diseases caused by a bacterial infection namely Mycobacterium tuberculosis. The radiologists spend more time for detecting the TB when analysing with traditional methods. However, manual detection of TB is time-consuming and challenging task in the current scenario. To overcome these challenges, a novel deep learning-based TOUR-STAR model has been proposed for classifying the TB into three classes. Initially, the chest x-ray (CXR) images are gathered from the Pad Chest dataset and the collected images are pre-processed by Gaussian filter for reducing the noises and smoothen the edges. The Dual Attention U-Network is used to segment the liver region separately from the pre-processed CXR images. Finally, the segmented images are fed into spiking dilated convolutional neural network (SDCNN) to classify the TB into tri classes such as normal controls (NC), Pulmonary Tuberculosis (PT) and Miliary Tuberculosis (MT). The proposed TOUR-STAR model is evaluated based on its f1 score, precision, specificity, recall and accuracy. The classification accuracy of 98.90% for the proposed SDCNN are highly reliable for Pad Chest dataset. The proposed SDCNN outperforms ResNet, AlexNet, RegNet, and GoogleNet by 5.30%, 1.63%, 4.49% and 0.83% in terms of overall accuracy range. The suggested TOUR-STAR model achieves the overall accuracy by 0.20%, 0.76% and 4.14% comparing to the existing method such as CNN, DNN and EfficientNet.