Narrow Results

Accurate and automated lung segmentation in high-resolution computed tomography (HRCT) is highly challenged by the presence of pathologies affecting lung parenchyma appearance and borders.

The algorithm presented employs an anatomical model-driven approach and systematic incremental knowledge acquisition to produce coarse lung delineation, used as initialization for the graph-cut algorithm. The proposed method is evaluated on a 49 HRCT cases dataset including various lung disease patterns. The accuracy of the method is assessed using dice similarity coefficient (DSC) and shape differentiation metrics (d_mean, d_rms), by comparing the outputs of automatic lung segmentations and manual ones.

The proposed automatic method demonstrates high segmentation accuracy (DSC = 96.64%, d_mean = 1.75 mm, d_rms = 3.27 mm) with low variation that depends on the lung disease pattern. It also presents good improvement over the initial lung segmentation (ΔDSC = 4.74%, Δd_mean = -3.67 mm, Δd_rms = -6.25 mm), including impressive amelioration (maximum values of ΔDSC = 58.22% and Δd_mean = -78.66 mm) when the anatomy-driven algorithm reaches its limit.

Segmentation evaluation shows that the method can accurately segment lungs even in the presence of disease patterns, with some limitations in the apices and bases of lungs. Therefore, the developed automatic segmentation method is a good candidate for the first stage of a computer-aided diagnosis system for diffuse lung diseases.