Accurate annotation of acid-fast bacilli (AFB) in images is essential for developing a reliable machine learning models for the diagnosis of extra-pulmonary tuberculosis (EP-TB). However manual annotation is time-consuming labour-intensive and prone to inter-observer variability limiting scalability and reproducibility. This study reports on the development and validation of an automated annotation tool to label AFB in whole-slide images of tissue biopsies for EP-TB detection. During inference the models generate explainable auto-annotations that support both diagnostic assistance and the creation of additional annotated datasets for further model training. The tool was evaluated using metrics such as sensitivity specificity annotation efficiency and usability. Results demonstrate improved annotation consistency and reduced annotation time. This study addresses a critical challenge in medical AI development: generating high quality annotations and curated datasets for pathology images.