3D Segmentation of Whole Lung and Metastatic Lung Nodules Using Adaptive Region Growing and Shape-based Morphology.

OBJECTIVE

Early diagnosis of primary and metastatic lung nodules is critical for effective therapeutic planning. Manual delineation of lung nodules is not time-efficient and is prone to human error as well as interobserver and intraobserver variability. This study aimed to address the unmet need for an open-source computer-aided detection (CAD) system for 3D segmentation of lung and metastatic lung nodules along with radiomic feature extraction.

METHODS

The proposed adaptive region-growing-based lung nodule segmentation (RGLNS) tool was developed in-house, requiring only manual input to select a seed point within the nodule on computed tomography (CT) images. A total of 230 CT scans from 100 patients with sarcomas were screened. Lung nodules were present in 200 CT scans, which were further analyzed. The accuracy of the lung and nodule segmentation was evaluated qualitatively using a 5-point Likert scale (uninterpretable: 1; poor: 2; fair: 3; good: 4; excellent: 5) and quantitatively using the Dice coefficient and Jaccard index.

RESULTS

A total of 200 CT scans comprising 12,000 CT slices were analyzed, among which 786 lung nodules were identified. Quantitative lung segmentation accuracies (n=2400 slices) yielded a Dice coefficient of 0.92±0.06 and a Jaccard index of 0.85±0.05. Qualitative scores (n=9600 slices) for lung boundary correction (4.56±1.18) and inclusion of pulmonary vessels (4.75±0.72) were rated as good to excellent. Quantitative nodule segmentation (n=142 nodules) accuracies were as follows: dice coefficient=0.92±0.03, 0.88±0.04, 0.86±0.03, 0.85±0.03, 084±0.04 and Jaccard index=0.84±0.03, 0.81±0.04, 0.78±0.04, 0.78±0.02, 0.76±0.04 for solitary (n=73), juxtapleural (n=32), juxtavascular (n=28), fissure-attached (n=6), and ground-glass (n=6) nodules, respectively. Qualitative scores (n=644 nodules) for nodule-boundary were good to excellent [solitary (n=342): 4.97±0.15; juxtapleural (n=155): 4.45±0.60; juxtavascular (n=127): 4.40±0.65; fissure-attached (n=9): 4.40±0.70; ground-glass (n=11): 4.25±0.75] and for exclusion of pulmonary vessels/pleura from nodules were good [juxtapleural (n=155): 4.10±0.66; juxtavascular (n=127): 4.08±0.64; fissure-attached (n=9): 4.30±0.67].

CONCLUSIONS

The proposed semiautomated CAD system, RGLNS, requiring minimal manual input, demonstrated robust, and promising segmentation results for the whole lung and various types of metastatic lung nodules.