Semi-automated Lung Field Segmentation in Scoliosis Radiographs: An Exploratory Study

Abstract

Scoliosis has a detrimental effect on lung function. Since spine radiographs are commonly used for monitoring the progress of the disorder, semi-automated lung field segmentation from scoliosis radiographs is a primary step in further automation and processing. Existing lung field segmentation algorithms have been developed specifically for chest radiographs, which differ from spine radiographs in imaging aspects and appearance. The present work uses intensity profile processing followed by the optimization of a flexible polynomial template developed without prior training data. Right and left lungs are processed separately due to the presence of the cardiac shadow and differences in lung shape. Intensity profile processing is based on identification of the maxima and minima in segments of the horizontal profiles or their derivatives. The polynomial template represents an implicit shape model defined separately for each lung using three initial points chosen by the user. Thus, a unique template is generated for each case. The algorithm was tested on six sets of patient radiographs varying in severity of scoliosis and image quality. The proposed method successfully detects overlapping regions of the lung and the spine and has high accuracy even for severely deformed lungs.