Computer vision techniques, particularly those employing Artificial Intelligence (AI) and Machine Learning (ML), have shown promise in the analysis of tuberculosis (TB) using digitized human tissue specimens. Our systematic review identified two prominent studies that applied deep learning frameworks to detect mycobacteria in digital pathology slides. Despite their reported successes in other contexts, our collaborative efforts to replicate and extend these models yielded unexpected results, prompting a thorough investigation into the potential causes behind their underperformance. This study comparatively analyses the deployment of two prominent models to the task of image-based TB detection. The research reveals that key factors contributing to underperformance include dataset limitations, model architecture complexities, annotation quality issues, transferability constraints, preprocessing discrepancies, and clinical relevance variations. While deep learning computer vision models hold significant potential for enhancing TB diagnosis through the analysis of human tissue specimens, our collaborative testing of two prominent models revealed several challenges that may hinder their performance in real-world applications. Addressing these challenges will be crucial for advancing the reliability and effectiveness of such models in clinical practice.