Domain Generalization of Tuberculosis Detection in Chest X-Rays through MixStyle and Multi-Layer Augmentation

In this study, we enhance tuberculosis (TB) detection in chest X-ray (CXR) datasets by combining domain generalization (DG) methods to simulate new domain styles. Tuberculosis remains one of the deadliest infectious diseases, yet millions of cases are undetected due to severe shortages of radiologists and diagnostic capacity, particularly in low-resource regions in Africa. In this regard, CXR analysis with Computer-Aided Detection (CAD) is a promising approach for TB detection in resource-constrained settings; however, performance often degrades under domain shifts such as differences in medical devices, datasets, or geographic locations. We thereby present a fine-tuned Convolutional Neural Network (CNN) baseline model that pairs MixStyle (feature-level style mixing) and AugMix (image-level robust augmentation). We fine-tune a DenseNet-121 model trained on publicly available TB chest X-ray datasets from NIAID TB using transfer learning. Our proposed model integrates DG techniques during training, extrapolating the domain to different variations and settings. This comparative analysis between the two models is evaluated on two fully unseen TB cohorts from Pakistan and the Shenzhen dataset. Relative to the traditional baseline, our approach improves accuracy from 66% to 89% on the Shenzhen dataset and from 71% to 78% on the Pakistani dataset, with accompanying gains in recall and precision. These results demonstrate that amalgamating DG techniques can substantially enhance cross-site generalization for TB classification across domains. Our findings suggest a feasible pathway for deploying reliable CAD systems in new geographic regions for patient triage without requiring additional labeled data from those specific locations, a major barrier to scaling AI-based TB screening in global health settings.