Lung Segmentation From Digital Radiography Chest Images

In the field of medical diagnosis a wide variety of imaging techniques are currently available,such as X-ray imaging, computed tomography (CT) and magnetic resonance imaging (MRI).Although the last two are more precise and more sensitive techniques, the chest radiography isstill by far the most common type of procedure for the medical examination and screening of thelungs, due to its less radiation dose and low cost [1]. The advent of digital thorax units anddigital radiology departments with picture archiving communication systems (PACS) makes itpossible to use computerized methods for the analysis of chest radiographs at a routine basis.The use of image processing techniques and computer aided diagnosis (CAD) systems hasproved to be effective for the improvement of radiologists’ detection accuracy for lung nodules inchest radiographs [2].The automatic detection of lung nodules, and in general the detection and characterization ofthe interstitial diseases in diagnostic chest radiology, are only interested in the information insidethe lung regions. In these cases, the lung segmentation becomes the first and critical procedure forthe computerized image analysis on the chest radiographs. However, the digital chest radiographsare characterized with variation in intensity contrast and non-uniform intensity background.Moreover, the complexity and overlap of anatomical structures in chest radiographs, introducemore difficulties and challenges for accurate lung segmentation.Recently, the problem of segmenting lung fields has attracted considerable attention withencouraging results. However, low-level segmentation methods that use only local intensitycriteria, such as thresholding, region growing and edge detection, are insufficient for accuratelung region delineation in digital chest radiographs. More powerful model based strategies arerequired that rely on a priori information about the shape of the object to guide the segmentationprocess.In this paper, a customized active shape model (ASM) to extract lungs from radiographychest images was proposed and validated. Firstly, the average ASM, gray-scale projection andaffine registration were employed to attain the initial lung contours. Secondly, a new objectivefunction with constraints of distance and edge was proposed to push the vertices of ASM to thereal lung edge, and to pull the vertices out of the stomach gas regions. Finally, multi-resolutionrepresentation and optimization were employed to enhance speed and avoid local optimization.Experimental results on a public database of247images showed that the proposed algorithmcould achieve an average accuracy of94.7%, which is4.4%better than the traditional ASM and2.7%better than the ASM with local invariant features. The customized active shape modelprovides a fast, effective, and model-based method for lung region segmentation in chestradiographs. The computerized techniques of lung segmentation that have been developed maybe incorporated into future CAD schemes and may potentially aid radiologists in theinterpretation and assessment of digital chest radiographs.