Evaluation of an automated lung segmentation method using an iterative thresholding and processing technique

Purpose: We previously reported automatic lung segmentation of CT scans using an atlas-based approach with multiple best matches. The objective of the current work is to evaluate an alternative method of lung segmentation which performs iterative Hounsfield Unit (HU) thresholding and processing steps to produce lung volumes-of-interest (VOIs).

Methods: The method described in the current work was applied to each CT scan in a 17-subject dataset. This method generates an initial VOI within the lungs and an initial trachea and bronchus VOI using HU thresholding and a sequence of other processing steps including VOI expansion and contraction, volume thresholding, hole-filling, and smoothing. To refine these initial VOIs, they are iteratively updated by expanding the current lung VOI by 5 voxels in each direction and then processing with a similar sequence of steps. The lungs are then separated into left and right components using a simple geometric split followed by additional processing steps designed to correct any regions which were falsely classified as right or left. Each scan had carefully hand-drawn lung VOIs available to use as reference standards for that subject. The Dice Similarity Coefficient (DSC) and Hausdorff mean distance were computed for each VOI resulting from this method compared to its corresponding reference standard VOI.

Results: Average DSCs achieved were 0.963 ± 0.012 (with a range of 0.936 to 0.977) and 0.958 ± 0.014 (with a range of 0.923 to 0.976) for the right and left lungs respectively, while the average Hausdorff mean distance across subjects was 0.859 ± 0.231 mm and 0.920 ± 0.266, respectively. These DSCs are comparable to what was determined from the atlas segmentation previously reported (0.955 and 0.956 for the right and left lungs respectively).

Conclusion: An iterative thresholding method for automatic lung segmentation has been described, capable of producing lung VOIs with DSCs comparable to the results of atlas segmentation with multiple best matches. This permits individuals without access to atlas segmentation methods to accurately and automatically generate lung VOIs. Reference: Pirozzi S, Horvot M, Piper J, Nelson AS. Atlas-based Segmentation: Evaluation of a Multi-Atlas Approach for Lung Cancer. Med Phys 2012; 39:3677.