Tuberculosis Lesion Segmentation Improvement in X-Ray Images Using Contextual Background Label

Pulmonary tuberculosis (PTB) is a serious, potentially fatal, infectious disease that primarily affects the lungs, and poses a significant threat to public health. To detect PTB at an early stage by screening chest X-Ray (CXR) images for tuberculosis (TB) lesions, we propose a semantic segmentation scheme that uses a deep learning algorithm. However, this scheme requires high-quality training data. To improve the TB-lesion segmentation model performance, a contextual background label process was designed for decomposing the heterogeneous CXR image background. From the designed process, five background subclasses consisting of: lung, mediastinum, body, doc, and background provided to modify the ground truth data for use in training the segmentation models, which was designed in four models to assess the performance improvement of the proposed scheme. The experimental results confirmed the applicability of the designed schemes. The TB-lesion segmentation models demonstrated improvements in terms of reduced false positives and better visualizations of the shape and location of TB lesions than the visualized approximation from the classification methods. The proposed model demonstrated highest scores of 88.68% on Dice, 83.55% on Intersection over Union, and 98.64% on precision for TB-lesion detection. The proposed models were externally validated to demonstrate their generalizability. They returned sensitivity, specificity and accuracy scores of 89.00%, 95.00% and 90.00%.