VAE-AL-UNet: Efficient Lung Segmentation in Chest X-Ray Images Using Variational Autoencoder and Active Learning

Tuberculosis (TB) remains a significant global public health concern as a highly infectious disease. One of the most effective ways of diagnosing TB is through chest X-ray (CXR) imaging. However, accurately segmenting lungs in CXR images is challenging due to image variations and disease complexity. Manually annotating many CXR images is a time-consuming and expensive process. This has led to a scarcity of labeled data, making it difficult to train accurate deep learning models for lung segmentation. To address these challenges, we propose a novel approach to automate the segmentation of lungs in CXR images. First, we trained a variational autoencoder (VAE) on an unlabeled dataset to learn the important features and extract the reconstructed images. This allowed us to capture underlying patterns and structures in the data. In the second step, we used the U-Net model for segmentation, which takes the original images along with the reconstructed images generated by the VAE to guide the segmentation process. In the other hand, Active Learning (AL) is used to select the most informative samples for annotation by using a two-step query method that involves calculating sample complexity and potential value. Sample complexity was calculated using a fusion of multiple estimate models, which allowed us to estimate the amount of data needed to achieve a certain level of performance. Potential value was defined as the activation value of a sample and was used to guide the interactive query process. Experimental results demonstrated that our method achieved an average Dice score of 95% and an Intersection over Union (IoU) of 92% on the testing set, indicating a high level of accuracy in lung segmentation, while minimizing the number of labeled examples needed to just 10% of the dataset. Our approach obtains state-of-the-art performance on two public CXR datasets. Overall, our proposed method is a promising approach for improving the accuracy and efficiency of lung segmentation in CXR images.