Clinically interpretable imaging biomarker discovery in BCG-vaccinated mycobacterium tuberculosis-infected diversity outbred mice using deep learning

Tuberculosis (TB), caused by <i>Mycobacterium tuberculosis</i> (M.tb), poses a substantial global health challenge, affecting over 2 billion individuals worldwide. Although the approved vaccine against TB, <i>M. bovis</i> Bacille Calmette-Guerin (BCG), successfully prevents TB meningitis and systemic TB in children, efficacy against pulmonary TB in adults is highly variable, ranging from 0&percnt; to 80&percnt; in epidemiological studies. As it can be challenging to identify correlates or biomarkers of successful BCG vaccination against pulmonary TB, we sought to develop a deep learning approach to distinguish BCG-vaccinated mice from non-vaccinated mice using hematoxylin and eosin (H&amp;E) stained lung sections from M.tb-infected Diversity Outbred mice and to identify granuloma features that may be unique to BCG vaccination. Upon analyzing the attention maps, we found that the dominant cellular features of granulomas from lungs of BCG-vaccinated mice included clusters of lymphocytes near macrophages within granulomas, and in contrast, the dominant features of lung granulomas from non-vaccinated mice included macrophages, foamy macrophages, and extracellular fibrillary materials. These results show the feasibility of our approach to identify imaging biomarkers using H&amp;E images.