Effective diagnosis of tuberculosis plays a pivotal role in preventing its spread, given that it is one of the most prominent causes of death. This study provides a comprehensive and critical analysis of two prevalent pre-trained object detection models, in the context of detecting the microscopic bacteria present in sputum smear images. The performance of YOLO family of models for object detection tasks and the FasterRCNN ResNet50 is evaluated using a transfer learning approach. Model evaluation is based on mean average precision, which is a widely accepted metric to monitor the performance of models in object detection tasks. YOLOv8 had better mean average precision of 0.86 compared to other YOLO models and FasterRCNN ResNet50. Our experiments confirmed YOLOv8 model has faster and more efficient detection of tuberculosis from microscopic image fields. Furthermore, we incorporated Layer-wise Relevance Propagation, a prevalent Explainable Artificial Intelligence method used for exploring the interpretability of bacilli detection on sputum smear images. This integration of Explainable Artificial Intelligence ensures transparency and instills confidence in model performance by unveiling model decisions and the relevance of every voxel in the image. This automated detection system can assist clinicians in accurately predicting the disease in a shorter timeframe, ultimately enhancing clinical outcomes.