Deep architectures for high-resolution multi-organ chest X-ray image segmentation

Abstract

Chest X-ray images (CXRs) are the most common radiological examination tool for screening and diagnosis of cardiac and pulmonary diseases. The automatic segmentation of anatomical structures in CXRs is critical for many clinical applications. However, existing deep models work on severely down-sampled images (commonly 256×256\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$256\\times 256$$\\end{document} pixels), reducing the quality of the contours of the resulting segmentation and negatively affecting the possibilities of such methods to be effectively used in a real environment. In this paper, we study multi-organ (clavicles, lungs, and hearts) segmentation, one of the most important problems in semantic understanding of CXRs. We completely avoid down-sampling in images up to 1024×1024\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$1024\\times 1024$$\\end{document} (as in the JSRT dataset), and we diminish its impact in higher resolutions via network architecture simplification without a significant loss in the accuracy. To do so, we propose four different convolutional models by introducing structural changes to the baselines employed (U-Net and InvertedNet) as well as by integrating several techniques barely used by CXRs segmentation algorithms, such as instance normalization and atrous convolution. We also compare single-class and multi-class strategies to elucidate which approach is the most convenient for this problem. Our best proposal, X-Net+, outperforms nine state-of-the-art methods on clavicles and lungs obtaining a Dice similarity coefficient of 0.938 and 0.978, respectively, employing a tenfold cross-validation protocol. The same architecture yields comparable results to the state of the art in heart segmentation with a Dice value of 0.938. Finally, its reduced version, RX-Net+, obtains similar results but with a significant reduction in memory usage and training time.