A stacking-based ensemble learning approach for lung segmentation in multi-center chest X-ray images

Automatic segmentation of lung fields in chest X-ray (CXR) images is an essential part of the diagnostic process. However, the lungs' diverse shapes and irregular appearance caused by disease make this task challenging. Most previous studies have relied on individual networks, each with its own strengths and weaknesses. In this study, a stacking-based ensemble learning approach incorporating multiple individual networks was developed to leverage their complementary strengths and improve these challenges. To evaluate the generalizability and reproducibility of the results, three public datasets, namely Montgomery, Shenzhen, and JSRT, were utilized. The outputs of the three models, FAT-Net, UNet3+, and DeepLabv3+, on the validation data were used as inputs to a convolutional neural network acting as a meta-learner. Finally, the performance of the proposed method was evaluated on the test data. The proposed method achieved improved performance across the evaluated metrics, with the two key segmentation metrics, IoU and Dice score, reaching 94.49% and 97.13%, respectively. These results indicate greater overlap between the predicted and ground-truth masks and more accurate segmentation. The robustness of these results was further confirmed through five-fold cross-validation, in which the proposed method consistently outperformed the individual models. Based on the results obtained and comparisons with recent studies, stacking-based ensemble learning, by integrating the strengths of individual models, can achieve more reliable results in lung field segmentation.