Repeatability of 18F-FDG PET Radiomic Features in Cervical Cancer

John P. Crandall; Tyler J. Fraum; MinYoung Lee; Linda Jiang; Perry Grigsby; Richard L. Wahl

doi:10.2967/jnumed.120.247999

Article Figures & Data

Figures

Tables
Additional Files

Download figure
Open in new tab
Download powerpoint
FIGURE 1.
Representative images. Tumors were manually delineated to generate VOI_WT (baseline image [A] and repeat image [B]). VOI₄₀ was created by removing all voxels with SUVs ≤ 40% of SUV_max of VOI_WT (baseline image [C] and repeat image [D]).
Download figure
Open in new tab
Download powerpoint
FIGURE 2.
Repeatable radiomic feature groups. Box plots show repeatability ranges for all features within specified feature group. ISR = isotropic voxels; NSR = no spatial resampling.
Download figure
Open in new tab
Download powerpoint
FIGURE 3.
Unrepeatable radiomic feature groups. ISR = isotropic voxels; NSR = no spatial resampling.
Download figure
Open in new tab
Download powerpoint
FIGURE 4.
Bland–Altman plots showing effect on radiomic feature repeatability of changes in reconstruction (PET_PSF CCC subtracted from PET_OSEM CCC [A]), segmentation (VOI_WT CCC subtracted from VOI₄₀ CCC [B]), spatial resampling (no-spatial-resampling CCC subtracted from resampling to 3-mm isotropic voxel CCC [C]), and SUV discretization (32-bin CCC subtracted from 64-bin CCC, 32-bin CCC subtracted from 128-bin CCC, and 64-bin CCC subtracted from 128-bin CCC [D]). Plots were generated by calculating mean of, and difference between, corresponding radiomic feature CCC values of each method.
Download figure
Open in new tab
Download powerpoint
FIGURE 5.
Bland–Altman plots showing how changes in reconstruction (A), segmentation (B), spatial resampling (C), and SUV discretization (D) affected repeatability of radiomic feature groups.

View popup

TABLE 1

Radiomic Feature Groups

Feature group	Features (n)	Radiomic feature names
Standard intensity	5	SUV_max, SUV_mean, SUV_peak, SUV_SD, total-lesion glycolysis
Nonstandard intensity	4	Skewness, kurtosis, entropy, uniformity
Shape	3	Metabolic tumor volume, sphericity, compacity
GLCM	6	Homogeneity, energy, contrast, correlation, entropy, dissimilarity
GLRLM	11	Short- and long-run emphasis, low and high gray-level run emphasis, short-run low and high gray-level emphasis, long-run low and high gray-level emphasis, gray-level nonuniformity, run length nonuniformity, run percentage
NGLDM	3	Coarseness, contrast, busyness
GLZLM	11	Short- and long-zone emphasis, low and high gray-level zone emphasis, short-zone low and high gray-level emphasis, long-zone low and high gray-level emphasis, gray-level nonuniformity, zone length nonuniformity, zone percentage

View popup
TABLE 2
Patient Characteristics
Characteristic Value
Age (y) 46.8 ± 11.0
Height (m) 1.66 ± 0.06
Weight (kg) 77.8 ± 17.3
Qualitative data are numbers followed by percentages in parentheses; continuous data are mean ± SD.

View popup

TABLE 3

Comparison of Radiomic Feature Repeatability Between PET Image Reconstruction Methods

Feature group	Feature	Mean PET_PSF CCC	Mean PET_OSEM CCC	P
Standard intensity	SUV_mean	0.82	0.86	<0.001^*
	SUV_SD	0.89	0.83	0.066
	SUV_max	0.88	0.86	<0.001^*
	SUV_peak	0.78	0.82	<0.001^*
	Total-lesion glycolysis	0.97	0.98	<0.001^*
Nonstandard intensity	Skewness	0.85	0.92	0.012
	Kurtosis	0.74	0.83	0.003
	Entropy	0.69	0.80	0.040
	Energy	0.70	0.80	0.003
Shape	Metabolic tumor volume	0.99	1.00	0.313
	Sphericity	0.88	0.88	0.016
	Compacity	0.99	1.00	0.313
GLCM	Homogeneity	0.89	0.97	<0.001^*
	Energy	0.89	0.94	0.110
	Contrast	0.90	0.95	0.176
	Correlation	0.93	0.91	<0.001^*
	Entropy	0.93	0.96	0.266
	Dissimilarity	0.91	0.96	0.176
GLRLM	Short-run emphasis	0.89	0.96	0.007
	Long-run emphasis	0.89	0.97	0.002
	Low gray-level run emphasis	0.52	0.58	0.176
	High gray-level run emphasis	0.62	0.78	<0.001^*
	Short-run low gray-level emphasis	0.52	0.59	0.176
	Short-run high gray-level emphasis	0.62	0.78	<0.001^*
	Long-run low gray-level emphasis	0.55	0.58	0.569
	Long-run high gray-level emphasis	0.71	0.82	0.002
	Gray-level nonuniformity	0.99	1.00	<0.001^*
	Run length nonuniformity	0.99	0.99	0.036
	Run percentage	0.89	0.96	0.009
NGLDM	Coarseness	0.96	0.97	0.493
	Contrast	0.88	0.89	0.522
	Busyness	0.93	0.92	0.092
GLZLM	Short-zone emphasis	0.54	0.74	0.005
	Long-zone emphasis	0.92	0.97	0.036
	Low gray-level zone emphasis	0.43	0.56	0.176
	High gray-level zone emphasis	0.54	0.73	<0.001^*
	Short-zone low gray-level emphasis	0.44	0.53	0.447
	Short-zone high gray-level emphasis	0.59	0.77	<0.001^*
	Long-zone low gray-level emphasis	0.55	0.77	<0.001^*
	Long-zone high gray-level emphasis	0.97	0.98	0.970
	Gray-level nonuniformity	0.98	0.99	0.009
	Zone length nonuniformity	0.94	0.98	0.003
	Zone percentage	0.83	0.94	0.007

↵* Significant P value, based on paired Wilcoxon signed-rank test. After correcting for multiple comparisons, P value of less than 0.001 was considered significant.

View popup

TABLE 4

Groupwise Comparison of Repeatability of Radiomic Features Extracted from PET_OSEM Images Using 2 Different Segmentation Methods

Feature group	Mean VOI_WT CCC	Mean VOI₄₀ CCC	P
Standard intensity	0.88	0.86	0.057
Nonstandard intensity	0.82	0.86	0.010
Shape	0.95	0.97	0.140
GLCM	0.93	0.96	<0.001^*
GLRLM	0.70	0.94	<0.001^*
NGLDM	0.91	0.94	0.023
GLZLM	0.73	0.90	<0.001^*

↵* Significant P value, based on paired t test. After correcting for multiple comparisons, P value of less than 0.004 was considered significant.
Mean value calculation includes CCC values from both resampling methods and each SUV bin level (i.e., 6 CCC values per feature per segmentation).

View popup

TABLE 5

Groupwise Comparison of Repeatability of Radiomic Features Extracted from PET_PSF Images Using 2 Different Segmentation Methods

Feature group	Mean VOI_WT CCC	Mean VOI₄₀ CCC	P
Standard intensity	0.87	0.87	0.984
Nonstandard intensity	0.75	0.74	0.784
Shape	0.95	0.97	0.015
GLCM	0.87	0.95	<0.001^*
GLRLM	0.68	0.81	<0.001^*
NGLDM	0.88	0.96	<0.001^*
GLZLM	0.68	0.73	0.139

↵* Significant P value, based on paired t test. After correcting for multiple comparisons, P value of less than 0.004 was considered significant.
Mean value calculation includes CCC values from both resampling methods and each SUV bin level (i.e., 6 CCC values per feature per segmentation).

View popup

TABLE 6

Groupwise Comparison of Repeatability of Radiomic Features Extracted from PET_OSEM Images Using 2 Different Spatial Resampling Methods

Feature group	Mean VOI_WT CCC			Mean VOI₄₀ CCC
	NSR	ISR	P	NSR	ISR	P
Standard intensity	0.88	0.88	0.857	0.85	0.87	0.021
Nonstandard intensity	0.82	0.82	0.970	0.86	0.86	0.890
Shape	0.95	0.94	0.090	0.97	0.96	0.250
GLCM	0.93	0.94	0.333	0.97	0.95	0.068
GLRLM	0.69	0.71	0.005	0.95	0.92	0.001^*
NGLDM	0.91	0.91	0.928	0.93	0.95	0.031
GLZLM	0.69	0.76	<0.001^*	0.90	0.90	0.249

↵* Significant P value, based on paired Wilcoxon signed-rank test. After correcting for multiple comparisons, P value of less than 0.002 was considered significant.
NSR = no spatial resampling; ISR = resampling to 3-mm isotropic voxels.
Mean value calculation includes CCC values from each SUV bin level (i.e., 3 CCC values per feature per resampling method).

View popup

TABLE 7

Groupwise Comparison of Repeatability of Radiomic Features Extracted from PET_PSF Images Using 2 Different Spatial Resampling Methods

Feature group	Mean VOI_WT CCC			Mean VOI₄₀ CCC
	NSR	ISR	P	NSR	ISR	P
Standard intensity	0.86	0.87	0.670	0.86	0.88	0.376
Nonstandard intensity	0.73	0.76	0.156	0.65	0.83	<0.001^*
Shape	0.95	0.94	0.683	0.97	0.96	0.729
GLCM	0.88	0.86	0.259	0.95	0.95	0.490
GLRLM	0.66	0.70	<0.001^*	0.82	0.79	0.045
NGLDM	0.89	0.88	0.655	0.96	0.97	0.345
GLZLM	0.63	0.72	0.001^*	0.75	0.71	0.077

↵* Significant P value, based on paired Wilcoxon signed-rank test. After correcting for multiple comparisons, P value of less than 0.002 was considered significant.
NSR = no spatial resampling; ISR = resampling to 3-mm isotropic voxels.
Mean value calculation includes CCC values from each SUV bin level (i.e., 3 CCC values per feature per resampling method).