Improving lung region segmentation accuracy in chest X-ray images using a two-model deep learning ensemble approach

We propose a deep learning framework to improve segmentation accuracy of the lung region in Chest X-Ray (CXR) images. The proposed methodology implements a “divide and conquer” strategy where the original CXRs are subdivided into smaller image patches, segmented them individually, and then reassembled to achieve the complete segmentation. This approach ensembles two models, the first of which is a traditional Convolutional Neural Network (CNN) used to classify the image patches and subsequently merge them to obtain a pre-segmentation. The second model is a modified U-Net architecture to segment the patches and subsequently combines them to obtain another pre-segmented image. These two pre-segmented images are combined using a binary disjunction operation to get the initial segmentation, which is later post-processed to obtain the final segmentation. The post-processing steps consist of traditional image processing techniques such as erosion, dilation, connected component labeling, and region-filling algorithms. The robustness of the proposed methodology is demonstrated using two public (MC, JPCL) and one proprietary (The University of Texas Medical Branch - UTMB) datasets of CXR images. The proposed framework outperformed many state-of-the-arts competitions presented in the literature.