Transformer-Based Drug-Resistant TB Diagnosis from Chest X-Rays: Multi-Class Evaluation and Deployment Insights

Drug-resistant tuberculosis (DR-TB) is a major global health problem, especially in areas with inadequate resources and diagnostic facilities. Chest X-ray (CXR) imaging is widely accessible but relies on expert interpretation, a bottleneck in burdened TB settings. Addressing data scarcity and computational constraints dominant in such environments, this study presents a data-efficient deep learning framework that uses a Data-Efficient Image Transformer (DEiT) model to automate DR-TB classification from CXR images. DEiT is rigorously benchmarked against leading neural networks (DenseNet121, InceptionResNetV2, MobileNetNetV2) in a curated data set of 7,961 CXRs from 13 countries with a high prevalence of drug-resistant tuberculosis. The DEiT model demonstrated superior diagnostic performance, attaining an AUC of 0.800 (vs. 0.787–0.792 for CNNs), with balanced precision (0.826) and recall (0.828). Bootstrapped statistical validation confirmed DEiT’s performance advantage. Importantly, DEiT’s efficient fine-tuning strategy, freezing the transformer encoder and updating only 0.24\% of parameters (202K/86M), resulted in reduced trainable parameters and faster convergence. This design improves suitability for deployment in under-resourced healthcare settings where hardware limitations and small data\-sets prevail. The knowledge distillation structure of the model is responsible for its improved generalization since it reliably utilizes pre-trained representations in situations with little data. In an extended multi-class classification task involving DS-TB, MDR-TB, Non-MDR-TB, and XDR-TB, DEiT maintained consistent discriminative performance. Targeted augmentation and class-balanced training further improved the recognition of underrepresented subtypes, reducing diagnostic bias. These findings underscore the potential of data-efficient transformer-based models to overcome the limitations of conventional CNNs. With appropriate integration into clinical workflows and infrastructure, such models could enhance the scalability and precision of drug-resistant tuberculosis diagnosis.