Automatic Segmentation of Lung Fields in Chest Radiographs

OBJECTIVE Accurate segmentation of lung fields in chest radiographs (CXR) is very useful for automatic analysis of CXR. In this work, we propose to use dense matching of local features and label fusion to automatically segment the lung fields in CXR. METHODS For an input CXR, the dense Scale Invariant Feature Transform (SIFT) descriptors and raw image patches were extracted as the local features for each pixel. The nearest neighbors of the local features were then quickly searched by dense matching directly from the whole feature dataset of the reference images. The dense matching included three steps: limited random initialization, propagation of nearest neighbor field, and limited random search, with iteration of the last two steps for several times. The label image patches for each pixel were extracted according to the nearest neighbor field and weighted by the matching similarity. Finally, the weighted label patches were rearranged as the label class probability image of the input CXR, from which thresholds were obtained for segmentation of the lung fields. RESULTS The Jaccard index of the proposed method reached 95.5% on the public JSRT dataset. CONCLUSION A high accuracy and robustness can be obtained by adopting dense matching of local features and label fusion to segment the lung fields in CXR, and the result is better than that of current segmentation method.