Table 1 Publications reporting methodological investigations on texture analysis in NSCLC patients.

Cheng⁴⁸

R

56

Staging, I–III (only T)

Impact of respiration-averaged CT on PET texture parameters

Adaptive threshold, threshold uptake 45% of the SUV_max ^*

FOS/IVH = 3 SS = 1 GLCM = 4 GLRLM = 3 NGTDM = 4

Texture parameters obtained with helical and respiration-averaged PET/CT showed a high degree of agreement (SUV entropy and entropy had the lowest levels of variation)

Cui⁵⁰

n.r.

20

n.r.

Impact of the segmentation method on tumor volume estimation (validation of DM algorithm)

Automatic (DM), fuzzy C-means, threshold uptake 40% of the SUV_max, threshold uptake 50% of the SUV_max, tumor-customized downhill, watershed^§

FOS/IVH = 1 NGTDM = 1 Gr = 1

DM algorithm was able to segment the tumor (also when adjacent to mediastinum or chest wall) and outperformed the other lung segmentation methods in terms of overlapping measure

Cui⁵¹

n.r.

40

n.r.

Impact of the segmentation method on tumor volume estimation (validation of topo-poly algorithm)

Threshold uptake 40% of the SUV_max, threshold uptake 50% of the SUV_max, adaptive threshold, fuzzy C-means, tumor-customized downhill, random walks, high-order interactive learning segmentation, PET/CT tumor-background likelihood model, topo-poly^§

NGTDM = 1

Topo-poly algorithm was able to delineate tumor margins better than other methods

Dong⁴⁵

R

50

Staging, I–IV

Impact of the segmentation method on tumor volume estimation

Absolute SUV cut-off of 2.5, manual (2 observers), threshold at 40% of the SUV_max ^*

FOS/IVH = 1 SS = 1 GLCM = 1  + visual score

Intratumor heterogeneity significantly correlated with differences in the GTV definition (high heterogeneity corresponded to a larger GTV)

Gao⁵⁷

n.r.

132

Staging, I–III

Impact of computer-based algorithm on diagnosis of mediastinal lymph node metastases (validation of computer-based algorithm)

Manual^#

FOS/IVH = 3 GLCM = 5 + visual score

Diagnostic ability of computer-based algorithm and visual experience was similar

Hatt⁴⁴

n.r.

25, only 17 analyzed

Staging, Ib–IIIb

Impact of the segmentation method on the tumor volume estimation

Adaptive threshold, fully automatic method (FLAB), manual, threshold at 50% of the maximum^*

FOS/IVH = 1 SS = 1

All delineation methods except the manual one resulted in underevaluation of MTV. Anatomic tumor size and heterogeneity were correlated (larger lesions were more heterogeneous)

Hofheinz⁴⁹

n.r.

30

n.r.

Impact of the segmentation method on tumor volume estimation (validation of voxel-specific threshold algorithm)

Lesion-specific threshold, manual, voxel-specific threshold^*°

FOS/IVH = 2 SS = 1

Voxel-specific threshold method was able to reproduce tumor boundaries accurately, independent of the heterogeneity

Leijenaar⁴⁰

n.r.

11 (test-retest cohort) + 23 (inter-observer cohort)

Features’ test–retest reliability and interobserver stability among multiple tumor delineation methods

Manual (by 5 observers), threshold at 50% of the maximum

FOS/IVH = 54 SS = 8 GLCM = 22 GLRLM = 11 GLSZM = 11

The majority of features had high test–retest (71%) and interobserver (91%) stability in terms of ICC

Leijenaar⁵²

P

35

Staging, I–III

Comparison of different discretization methods for textural features

Manual (SUV discretization using a fixed bin size and a fixed number of bins)

GLCM = 22 GLRLM = 11 GLSZM = 11

SUV discretization had a crucial effect on textural features

Oliver⁴²

R

23

Sensitivity of texture features to tumor motion by comparison of static (3D) and respiratory-gated (4D) PET imaging

Adaptive threshold (background-adapted thresholding method)^*

FOS/IVH SS GLCM GLRLM (total 56)

Quantitative analysis using a 3D versus 4D acquisition provided notably different image feature values, mainly due to the impact of respiratory motion

Orlhac⁴⁶

P

24

Staging, III

Impact of the segmentation method on the tumor volume estimation

Threshold at 40% of the maximum, adaptive threshold^*°

FOS/IVH = 8 SS = 1 GLCM = 6 GLRLM = 11 GLSZM = 11 NGLDM = 3

IVH-based indices strongly depended on the tumor delineation method; 17/31 second- or high-order statistic features were robust with respect to tumor segmentation. Several texture indices included similar information. Some texture indices were highly correlated with MTV

Orlhac⁵³

R

48

Staging, I–III

Impact of resampling step on textural features and on the ability of textural features to reflect tissue-specific patterns of metabolic activity

Adaptive threshold (relative resampling approach and absolute resampling approach)^*°

FOS/IVH = 1 SS = 1 GLCM = 2 GLRLM = 3 GLSZM = 2

Textural features computed using an absolute resampling method varied as a function of the tissue type and cancer subtype more than when using the usual relative resampling approach

Tixier⁵⁵

P

20

Staging, I–II

Impact of static and parametric acquisition on PET features

Fully automatic method (FLAB)^*°

FOS/IVH = 2 SS = 3 GLCM = 3 GLSZM = 2

Compared with static SUV images, parametric images did not provide significant complementary information concerning heterogeneity quantification

van Velden⁴¹

P

11

Staging, IIIb–IV

Repeatability of texture features using different reconstruction settings and delineation methods

Threshold uptake 50% of the 3D SUV_peak on EANM-compliant (reconstruction method 1) and PSF-based (reconstruction method 2) images^°

FOS/IVH = 29 FF = 3 SS = 10 GLRLM = 22 GLCM = 44 L = 1 SA = 2

The majority of features had a high level of repeatability (ICC ≥ 0.90 for 63 features). Features were more sensitive to a change in delineation method (n = 25) than a change in reconstruction method (n = 3)

Yan⁴⁷

R

17

n.r., I–IV

Variability of PET textural features using different reconstruction methods, iteration numbers, and voxel size

Threshold uptake 40% of the SUV_max ^*°

FOS/IVH = 6 GLCM = 21 GLRLM = 11 GLSZM = 13 NGLDM = 5 NGTDM = 5

Image features had different sensitivities to reconstruction settings (entropy_Hist, difference entropy, inverse difference normalized, inverse difference moment normalized, low gray-level run emphasis, high gray-level run emphasis, and low gray-level zone emphasis were the most robust features; skewness, cluster shade, and zone percentage exhibited large variations)

Yip⁵⁶

R

26

Staging, n.r.

Sensitivity of texture features to tumor motion by comparing static (3D) and respiratory-gated (4D) PET imaging

Threshold uptake 40% of the SUV_max

GLCM = 1 GLRLM = 1 NGTDM = 4

4D-PET derived textures were less susceptible to tumor motion and may have greater prognostic value