Interpretation of Chest X-rays using Deep Learning Techniques for Early Diagnosis of Tuberculosis

Tuberculosis is a health challenge in Sri Lanka and a leading cause of mortality. Early diagnosis is necessary to prevent permanent lung damage and improve treatment outcomes. This research aims to develop a highly accurate and solid TB predictive model using chest X-rays with advanced deep-learning techniques. A key contribution of this study is to enhance the diagnostic capability of TB by introducing a novel dataset, "Non-TB," which exhibits radiological features similar to those of TB and other lung diseases. This study incorporates the Sri Lankan dataset, manually collected and reviewed by experts from local hospitals, alongside the publicly available Tuberculosis dataset for comprehensive analysis. Convolutional Neural Networks and advanced architectures like standard CNN approaches, DPN, ResNet50, and EfficientNetB7 were utilized for image segmentation and classification. The evaluation results demonstrated that the U-Net model achieved 99% accuracy in lung segmentation, while the ResNet50 model showed superior performance with the Sri Lankan dataset. Introducing the Non-TB class significantly improved diagnostic accuracy, with the multiclass model (R50SLTB02) achieving approximately 98% accuracy compared to the binary-class model (R50SLTB01), which gained 94.81%. This study increased the accuracy of diagnosis and reduced false positives, which are usually a significant challenge when other lung diseases present similarly in the X-ray. Therefore, the proposed model can seamlessly integrate into existing medical imaging systems, particularly in resource-limited settings, to facilitate early TB diagnosis. Future work will focus on expanding the Sri Lankan dataset and incorporating additional diagnostic modalities to improve diagnostic capabilities further.