Computer-Aided Diagnosis Systems for Lung Cancer: Challenges and Methodologies

Table 4

Studies on volumetric nodule segmentation reported from 2006 to 2010. See Table 3 for description of captions.

Study	Purpose	Type	Method	Database	Performance
El-Baz et al. [160]; Farag et al. [161]	General, cavity	Deformable model, 3D	Lesion boundary optimization by fitting a prior model with MRF and an appearance model with a bimodal LCDG	350 nodules with 3 to 30 mm of 29 patients; manual GTs by a radiologist	Segmentation error: min. 0.4%, max. 2.25%, mean 0.96%
van Ginneken [162]	General	Discriminative classification	Soft segmentation by supervised classifier. KNN regression of voxel-wise nodule probability with intensity features (gradient magnitude, Hessian eigenvalues, etc., over Gaussian scale-space)	LIDC1 dataset: 23 nodules with manual GTs	Average soft-overlap: 0.66±0.18 (0.52±0.25 by [73]); average percentage volume error: 23.7%±124.5% (by 510.8%±1577.7% [73])
Way et al. [163, 164]	General	Deformable model, 2D/3D	Successive 2D active contour with 3D gradient, 3D curvature, and mask energy terms with greedy optimization	LIDC1 dataset: 23 nodules with manual GTs	Average overlap: ranging between 0.07 and 0.63 across varying probabilistic GTs; median percentage volume error: 40%
Goodman et al. [165]	Juxtavascular	Watersheds	Watersheds segmentation followed by a model-based shape analysis to handle juxtapositions	50 nodules of 25 patients (<20 mm) with 17 irregular/spiculated margins,; 16 juxtapleural, 10 juxtavascular, and 2 GGO cases	Success rate: 97% over 450 measures (3 time-points by 3 observers)
Zhou et al. [166, 167]	GGO, juxtavascular	Probabilistic classification	Voxel-wise classification by comparing a nonparametric kernel density estimate of GGO intensity model with that of each local neighborhood by the Bhattacharya distance. Juxtavascular cases by eigen analysis of Hessian	10 GGO nodules	Only qualitative assessment
Yoo et al. [168]	GGO, juxtavascular	Deformable model, 3D	Asymmetric 3-phase deformable model of two LS functions	3 nodules	Only qualitative assessment
Wang et al. [169]	General	Dynamic programming, 3D	Transformation of 3D image to 2D polar-coordinate image by spiral scanning followed by 2D contour optimization by DP	LIDC1 dataset: 23 nodules with 4.0 to 33.6 mm diameter; LIDC2 dataset: 73 nodules with 3.8 to 30.2 mm diameter	Average overlap (LIDC1): 0.66 in [0.47, 0.89]; average overlap (LIDC2): 0.64 in [0.39, 0.87]
Nie et al. [170]	General	MS, 2D	MS clustering on a feature domain of convergence index by [171]	39 nodules with manual GTs	Average overlap: 0.83
Zheng et al. [172]; Zheng et al. [173]	General	Graph-cut, coupled segmentation-registration, 2D, automatic	2D graph-cut segmentation coupled with B-spline nonrigid lung registration [172]. Spatially coherent segmentation by solving MRF with graph-cut [173]	12 nodules with manual GTs	Mean percentage of the nodule volume variation: 0.8±0.6 in [172]
Browder et al. [174]	GGO, small, juxtavascular	Probabilistic classification	3-class (solid, non-solid, parenchyma) voxel-wise probabilistic classification with Gaussian intensity model. Bilateral filter by Tomasi used for noise removal. Juxtavascular cases by vessel removal filtering	ELCAP dataset: 75 cases with 5.6–17.5 mm in diameter; manual GTS by radiologists	Median growth consistency by geometric closeness metric: 1.87 (3.12 by radiologists)
Dehmeshki et al. [175]	Juxtavascular, juxtapleural	Region growing	Sphericity-oriented contrast-based region growing from an optimum seed point within a fuzzy connectivity map	815 nodules with 5–30 mm in diameter, 98 juxtapleural or juxtavascular cases	Success rate: 85–83%
Diciotti et al. [176]	Small, juxtavascular,	Semiautomatic, region growing	Target detection by LoG filtering followed by user selection. 3D region growing segmentation using a fusion-segregation criteria with geodesic distance	Phantom: 60 solid, juxtavascular, non-solid cases with 5.3–11 mm in diameter; ITALUNG dataset: 98 nodules; LIDC1 dataset: 23 nodules	Success rate: 86.3% (ITALUNG: 79.7% for juxtavascular); 83.3% (LIDC1: 75% for juxtavascular); volumetry RMS error: 1.0–6.6%
Kubota et al. [177, 178]	Small, juxtapleural, juxtavascular, solid, GGO	Region growing	Nodule enhancement by coupled competition-diffusion filtering. nodule core estimated as the maximum component of Euclidean distance map. Juxtapleural cases by estimating region core with centricity map. Segmentation and nodule extraction by region growing followed by convex hull	LIDC1: 23 nodules; LIDC2: 82 nodules; 820 nodules with manual diameter GTs	Average overlap (LIDC1): 0.69±0.18 (0.67±0.22 by [153], 0.57±0.20 by [73], 0.52±0.25 by [157]); average overlap (LIDC2): 0.59±0.19 (0.56±0.18 by [153], 0.45±0.21 by [157])
Zheng et al. [179]	GGO	Opacity map estimation, 2D	Thresholding opacity map estimated by solving a linear equations system constructed with the graph Laplacian	40 slices of 11 patients; manual GTs	Average shortest distance along contours: 4.2±4.9 pixels
Wang et al. [180]	General	Dynamic programming, 3D	Multidirection segmentation fusion by sequential dynamic 2D contouring ([145, 169]) applied to three orthogonal directions of a volume	LIDC1: 23 nodules for training; LIDC2: 64 nodules for testing	Overlap (LIDC1): mean 0.66, true-positive rate (TPR): 75%, false-positive rate (FPR): 15%; overlap (LIDC2): mean 0.58, true-positive rate (TPR): 71%, false-positive rate (FPR): 22%
Tao et al. [181]	GGO	Probabilistic classification	GGO nodule class-conditional probability map derived by an iterative LDA with GMMs of various intensity features.; Nodule segmentation by applying shape-prior probability mask	1100 nodules with 100 GGO nodules; 60 cases with manual GTs	Average overlap: 0.68; voxel-wise classification success rate: 92.28% overall; 89.87% GGO