A comparative analysis of tuberculosis-infected lung x-ray image segmentation: U-Net vs. U-Net++

Tuberculosis (TB) is a significant global health issue, causing a high number of deaths each year, with 1.6 million people dying from TB in 2021. Lung segmentation is a crucial step in TB detection and other lung disorder diagnoses. Deep learning, particularly convolutional neural networks (CNNs), has shown promise in accurately segmenting lungs in medical images like chest x-rays or CT scans. U-Net is an encoder-decoder CNN-based model with a U-shaped architecture used for image segmentation. It introduces skip pathways layer-wise between encoder-decoder pathways, for reducing sematic loss of information that happens while convolving the images in the encoder pathway. The U-Net++ on the other hand uses reengineered skip connections to further reduce the semantic information gap between all the layers of encoder and decoder. In this work, both U-Net and U-Net++ has been used for segmenting tuberculosis annonated lung images. The dataset contained 566 x-ray images. The results clearly showed that U-Net++ performed better than U-Net in terms of Dice index (14%), pixel accuracy (4.9%), and specificity (6.5%) while interestingly, U-Net fared better at specificity (3.2%).