This paper presents a novel deep learning-based framework, TB-CXRNet, for the accurate and reliable detection of tuberculosis (TB) and drug-resistant TB using chest X-ray (CXR) images. The framework aims to address the challenges of early and accurate TB diagnosis, particularly in the context of increasing drug-resistant TB cases. The study utilizes a large dataset of 40,000 CXR images, which includes 10,881 healthy, 24,119 non-TB, and 5,000 TB patients. The proposed method involves two main stages: first, classifying CXR images into healthy, non-TB, and TB categories using a pre-trained CheXNet model and a Self-MLP classifier; second, stratifying TB patients into drug-sensitive and drug-resistant groups using a stacking machine learning model. The stacking model combines the top three performing classifiers (Gradient boosting, Adaboost, and Logistic regression) to improve overall accuracy. The proposed approach achieves an accuracy of 93.32% for classifying CXR images into healthy, non-TB, and TB categories, and 87.48% for binary classification (drug-resistant vs drug-sensitive TB), and 79.59% for three-class classification (MDR, XDR, and sensitive TB). The study also introduces Score-CAM visualization techniques to enhance model interpretability, showing that the model learns from specific regions of the CXR images. The proposed framework can significantly aid in early detection and treatment of TB and drug-resistant TB, reducing disease complications and spread.This paper presents a novel deep learning-based framework, TB-CXRNet, for the accurate and reliable detection of tuberculosis (TB) and drug-resistant TB using chest X-ray (CXR) images. The framework aims to address the challenges of early and accurate TB diagnosis, particularly in the context of increasing drug-resistant TB cases. The study utilizes a large dataset of 40,000 CXR images, which includes 10,881 healthy, 24,119 non-TB, and 5,000 TB patients. The proposed method involves two main stages: first, classifying CXR images into healthy, non-TB, and TB categories using a pre-trained CheXNet model and a Self-MLP classifier; second, stratifying TB patients into drug-sensitive and drug-resistant groups using a stacking machine learning model. The stacking model combines the top three performing classifiers (Gradient boosting, Adaboost, and Logistic regression) to improve overall accuracy. The proposed approach achieves an accuracy of 93.32% for classifying CXR images into healthy, non-TB, and TB categories, and 87.48% for binary classification (drug-resistant vs drug-sensitive TB), and 79.59% for three-class classification (MDR, XDR, and sensitive TB). The study also introduces Score-CAM visualization techniques to enhance model interpretability, showing that the model learns from specific regions of the CXR images. The proposed framework can significantly aid in early detection and treatment of TB and drug-resistant TB, reducing disease complications and spread.