Table 2.

Radiomic articles on methodology, detailing effects of different acquisition and reconstruction parameters on shape features.

Author/title	Cancer site	Images used for analysis	Radiomics shape features	Other feature classes included	Results
Zhao, 201424 Exploring Variability in CT Characterization of Tumors: A Preliminary Phantom Study	Thorax phantoms with 22 lesions of varying sizes, shapes and densities	1.25, 2.5 and 5 mm slice thickness, Lung and standard reconstruction filters	Compactness, shape index 9 (proportion of the “spherical cap” of the nine types of shapes), fractal dimension, fractal lacunarity	First order statistics and texture features	All 14 features were significantly different between images with 1.25 and 5 mm slice thickness
Kalpathy-Cramer, 201625 Radiomics of Lung Nodules: A Multi-Institutional Study of Robustness and Agreement of Quantitative Imaging Features	Lung nodules	40 NSCLC and 12 phantoms with 9 different segmentations each	7 different centers with varying definitions and number of extracted features including the categories: global shape descriptors, local shape descriptors, margins	First order statistics and texture features	68% of the total 830 features (and 63% of shape features) exhibit stability to different segmentations with CCC ≥ 0.75
Lu, 201613 Assessing Agreement between Radiomic Features Computed for Multiple CT Imaging Settings	32 NSCLC patients (raw imaging data from RIDER dataset)	Varying slice thicknesses (1.25, 2.5 and 5 mm) and reconstruction filter (Lung [L] and Standard [S])	Compact-Factor, Eccentricity, Round-Factor (2D), Solidity (ratio of the object area over the area of the convex hull bounding the object), Shape Index features capturing the intuitive notion of ‘local surface shape’ of a 3D object (spherical cup, trough, rut, saddle rut, saddle, saddle ridge, ridge, dome, spherical cap)	First order statistics, texture and wavelet features	Hierarchical clustering grouped 89 features to 23 nonredundant groups. Majority of the shape-based features showed stability with average CCC values > 0.8 across all of the 15 inter-setting comparisons. Using the same reconstruction filter with either a 1.25 or a 2.5 mm slice thickness showed the best agreement
Desseroit, 201726 Reliability of PET/CT shape and heterogeneity features in functional and morphological components of NSCLC tumors: a repeatability analysis in a prospective multi-center cohort	Stage IIIB-IV NSCLC Merck MK-0646-008 (40 pts in 17 sites); ACRIN 6678 (34 pts in 14 sites) trials	71 primary tumors and 5 additional lesions	Four shape descriptors: sphericity, irregularity, major axis, 3D surface	First order statistics and texture features	Quantization/discretization was important in the reliability of features, with CT-based features more stable with fixed bin width. Morphological irregularity, sphericity and 3D surface were the most repeatable (Bland-Altman analysis of the differences between standard deviations of 3.3%, 10.0% and 11.6%, respectively)
Oliver, 201727 Sensitivity of Image Features to Noise in Conventional and Respiratory-Gated PET/CT Images of Lung Cancer: Uncorrelated Noise Effects	31 NSCLC patients	4 image sets per patient (original, low, medium, and high noise for 3D & 4D PET, 3D & 4D CT)	11 shape features: Volume, Surface area, Surface-to-volume, Sphericity, Compactness Spherical disproportion, Long axis, Short axis, Eccentricity, Convexity	22 first order, 26 GLCM, 11 GLRLM, and 11 GLSZM features	In both PET and CT, shape features exhibit the least change when uncorrelated noise is added (<13% average difference in CT)
Ul-Hassan, 201731 Intrinsic dependencies of CT radiomic features on voxel size and number of gray levels	ABS 3D printed phantoms, with a spherical contoured ROI of 4.2 cm3	116 CT scans, resampled to 1 × 1 × 2 mm3 voxel size	10 shape features: Convexity, Volume, Surface area, Surface-to-volume, Compactness, Long axis, Sphericity, Spherical disproportion, Short axis, Eccentricity	First order statistics(16), GLCM (24), GLZSM (11), fractal dimensions, texture and wavelet features	Shape features are robust, with eight out of the 10 having COVs < 50% with a negligible effect of resampling. The remaining two had diminished COV (<30%) after resampling

LoG: Laplacian of Gaussian; NSCLC: Non-small cell lung cancer; CCC: concordance correlation coefficient; GLCM: Gray-Level Co-occurrence Matrix; GLZSM: Gray-Level Size Zone Matrix; GLRLM: Gray-Level Run Length Matrix; SD: standard deviation; ACRIN: American College of Radiology Imaging Network; RIDER: Reference Image Database to Evaluate Therapy Response; ABS: Acrylonitrile Butadiene Styrene; COV: coefficient of variation.