Vision transformer-enabled chest CT analysis for enhanced tuberculosis diagnosis in clinical settings

Tuberculosis (TB) remains a leading cause of illness and mortality worldwide, particularly in resource-constrained regions where timely diagnosis is critical for effective treatment. Chest computed tomography (CT) offers detailed visualization of lung structures and can reveal subtle abnormalities such as cavitations, consolidations, and nodular lesions that are often missed in chest radiographs. However, the scarcity of annotated TB datasets and the heterogeneity of disease manifestations present major challenges for developing robust computer-aided diagnostic systems. To address this, we propose a Domain-Adaptive Pretraining and Fine-Tuning Vision Transformer (DAP-ViT) framework tailored for TB detection. The approach leverages self-supervised learning on a large, publicly available chest CT dataset to capture general radiological features and then applies adversarial domain adaptation to reduce the shift between source cancer images and TB-specific scans. Also, radiomics-guided fusion is employed to integrate handcrafted descriptors such as texture and shape, enhancing the model's sensitivity to clinically meaningful patterns. Fine-tuning is conducted on a curated TB CT dataset to optimize detection performance. Experimental results demonstrate that DAP-ViT achieves 95.0 % accuracy and 93.5 % recall, outperforming baseline convolutional networks by 7 % and 8.5 % respectively. The integration of radiomics improves interpretability by aligning model predictions with radiological findings, thus supporting clinical trust and adoption. By simplifying data requirements and improving generalization across domains, this framework has the potential to strengthen early TB diagnosis, reduce diagnostic delays, and support clinical decision-making in diverse healthcare settings.