Tuberculosis Chest X-ray Segmentation Using V-Net Deep Learning Algorithm

Tuberculosis (TB) has continued to be an ailment to the entire world, and proper diagnosis based on correct and unbiased treatment is vital in order to assimilate and alleviate this condition. Chest X-rays (CXRs) are commonly applied in screening of TB, but their manual interpretation is subjective and resource-laden. This work is based on determining the efficacy of V-Net deep learning architecture in automated TB lesions segmentation across CXR images. The model was trained with a curated and preprocessed dataset having binary masks defined the region in which TB is present. It was quantitatively assessed based on high segmentation accuracy, which was found with Dice coefficient, 0.932, Intersection over Union (IoU), 0.873, precision-0.911, recall 0.954 and overall accuracy 0.937. Training and validation accuracy and loss curves allowed estimating that learning is stable with little overfitting based on a confusion matrix that proved very high sensitivity and specificity. Qualitative results of the prediction of segmentation masks demonstrated reasonable anatomical consistency in the predicted results which closely conform with the clinical expectation especially in the upper lobes which are the common areas afflicted by TB. Excellent segmentation results confirm the reliability and interpretability ofV-Net as an effective means of TB lesions in chest radiographs, which can be used with great opportunities in automated screening systems, particularly at resource-limited locations.