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

Table 3

Studies on volumetric nodule segmentation reported from 1998 to 2005. Studies are ordered by their publication year. The purpose, type, and basic idea of each reported segmentation method are briefly described. Data and method used for validation of the proposed methods are also described.

Study	Purpose	Type	Method	Database	Validation and performance
Kawata et al. [137, 138]	Solitary, solid	Deformable model, 3D	Geometric deformation flow of 3D LS surface proposed by Caselles et al. [139]	62 nodules (47 malignant 15 benign) between 6 and 25 mm	Qualitative: correct segmentation of nodules with ill-defined surface; malignancy classification with two 3D surface characteristics
Yankelevitz et al. [140, 141]	Small, solitary, solid	Threshold (2D [141]/3D [140])	𝑘-means segmentation for automatic threshold estimation	Phantom (3.20 and 3.96 mm); in vivo: 13–15 nodules in repeat CTs	RMS error in volume measurement: ±3% (3D); volumetry: effective measurement of malignant growth of nodules as small as 5 mm (2D) with doubling time less than 177 days (3D)
Zhao et al. [142, 143] and Wiemker and Zwartkruis [144]	Small, juxtavascular	Threshold (2D [142]/3D [143])	Multicriterion automatic threshold estimation with average gradients along lesion contour and with boundary shape compactness, lesion segmentation by CCL and MOs, efficient average gradient computation [144]	9 nodules (<10 mm) with manual GT (2D)	Mean difference of 0.97 pixels was not statistically significant: 𝑃=0.90 (2D)
Xu et al. [145]	Juxtavascular, juxtapleural, calcification	Dynamic programing, 2D	2D contour optimization by DP. Calcification removal by EM classification of air, soft and calcified tissues. Semiautomatic contour correction by observers	4 nodules	Qualitative discussion only
Fetita et al. [76]	Juxtavascular, juxtapleural	Automatic, mathematical morphology	Gray-level MO with SMDC-connection cost. Juxtavascular cases by morphological dilation. Juxtapleural cases by global lung segmentation	300 nodules with 2–20 mm diameters of 10 patients	Detection performance: 98% sensitivity and 97% specificity for isolated and juxtavascular nodules; 90% sensitivity and 87% specificity for juxtapleural nodules
Ko et al. [146]	Small, solid/GGO	Threshold	Two-value thresholding with partial-volume correction based on CT intensity values	Phantom: 40 synthetic nodules (<5 mm, 20 solid and 20 GGO)	Average error in volume measurement: 2.1 mm^3
Kostis et al. [73, 147]	Small, juxtapleural, juxtavascular	Mathematical morphology	Isotropic resampling for partial-volume effect; binary segmentation by thresholding and CCL followed by vascular subtraction and pleural surface removal with iterative MOs	105 small nodules (<10 mm) of two time-points	Success rate: 80% for 21 juxtavascular cases; reproducibility study in measuring the percentage volume changes [147]
Okada et al. [148–151]	Small, juxtavascular, GGO	Robust anisotropic Gaussian fitting and mean shift (MS)	Robust anisotropic Gaussian intensity model fitting with MS segmentation in 4D spatiointensity domain	77 nodules of 3–25 mm diameters of 14 patients	Success rate: 89.6%; consistency: 1.12 voxel mean error for lesion center estimate when perturbing initialization
Kuhnigk et al. [152, 153]	Small, juxtavascular, juxtapleural	Automatic, region growing, and mathematical morphology	Region growing and CCL for initial segmentation. Juxtapleural and juxtavascular cases by convex hull and MOs. Volume estimation with partial-volume effect handling	Phantom: 31 nodules of various types; in vivo: 105 nodules with diameter larger than 4.6 mm of 16 patients	Success rate: 91.4%; inter-observer variability: 0.1% median error and 7.1% error at 95% limit; inter-scan variability: 4.7% median error and 26.9% error at 95% limit; volumetry median error with phantom: −3.1% for vascularized cases; −10.2% for juxtapleural cases
Mullally et al. [154]	Solitary, solid	Automatic, threshold	Automating the selection of VOI for thresholding-based segmentation methods by Zhao et al. [142, 143] and Ko et al. [146]	Phantom: 40 nodules (2.4 and 4.9 mm); in-vivo: 29 nodules in repeat CTs; manual GTs by a radiologist	Volume accuracy: 43% error for phantoms; 50% error for in vivo nodules
Shen et al. [155]	Juxtapleural	Surface analysis	Lung surface removal for juxtapleural nodule segmentation by local surface smoothing	20 juxtapleural nodules of a patient	Average RMS deviation from median by various click points: <2% except for one case; volumetry consistency: 60% of all varying click points leads to the same volume measure
Zhang et al. [97, 156]	GGO, juxtavascular	Probabilistic classification	MAP segmentation with a conditional distribution by a two-class GMM and with a priori by MRF. MAP optimization solved by iterated conditional modes. Juxtavascular cases by vessel segmentation. Conditional distribution adapted to each nodule to account for intensity offsets [156]	23 GGO nodules of 8 patients; manual GTs by two radiologists [156]	Success rate: 91.3%; consistency with 3 different clicks: 0.96±0.02 overlaps for all 21 successfully segmented cases; average overlap with GTs: 0.69±0.05; interobserver consistency: 0.73±0.04
Okada et al. [157]; Okada et al. [158, 159]	Small, juxtavascular, GGO, juxtapleural	Probabilistic classification [157] and mathematical morphology [158, 159]	Likelihood ratio test in spatiointensity joint domain after robust anisotropic Gaussian fitting by [151]. Juxtapleural cases by morphological opening and by prior constrained MS for rib bone suppression	1312 nodules of 39 patients; 123 true-negative cases included 108 juxtapleural cases	Success rate: 83.5% by [157]; 94.8% by [158, 159] overall; 71.5% for the juxtapleural/true-negative cases

GGO: ground-glass opacity (nonsolid and partially solid) nodules; LS: level sets; DP: dynamic programing; MO: morphological operations; CCL: connected-component labeling; EM: expectation-maximization; MAP: maximum a posteriori; MRF: Markov’s random fields; KNN: 𝑘-nearest neighbor; GMM: Gaussian mixture model; LDA: linear discriminant analysis; GT: segmentation ground truth.