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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Tuberculosis (TB) is a global health challenge requiring timely diagnosis, often hindered by limited annotated data for AI models.
• Chest computed tomography (CT) provides detailed lung visualization but faces challenges in AI development due to data scarcity and disease heterogeneity.

Purpose of the Study:

• To develop a robust computer-aided diagnostic system for TB detection using chest CT.
• To address data limitations and improve the generalization of AI models for TB diagnosis.

Main Methods:

• Proposed a Domain-Adaptive Pretraining and Fine-Tuning Vision Transformer (DAP-ViT) framework.
• Utilized self-supervised learning on public CT data and adversarial domain adaptation.
• Integrated radiomics features for enhanced pattern recognition and interpretability.

Main Results:

• DAP-ViT achieved 95.0% accuracy and 93.5% recall in TB detection.
• Outperformed baseline convolutional networks by significant margins (7% accuracy, 8.5% recall).
• Radiomics integration improved model interpretability and alignment with radiological findings.

Conclusions:

• The DAP-ViT framework effectively enhances TB detection from chest CT scans.
• The approach simplifies data requirements and improves model generalization across different domains.
• This AI tool supports early TB diagnosis, reduces delays, and aids clinical decision-making in various healthcare settings.