Adding the temporal domain to PET radiomic features

Wyanne A Noortman; Dennis Vriens; Cornelis H Slump; Johan Bussink; Tineke W H Meijer; Lioe-Fee de Geus-Oei; Floris H P van Velden

doi:10.1371/journal.pone.0239438

. 2020 Sep 23;15(9):e0239438. doi: 10.1371/journal.pone.0239438

Adding the temporal domain to PET radiomic features

Wyanne A Noortman ^1,^2,^*, Dennis Vriens ¹, Cornelis H Slump ³, Johan Bussink ⁴, Tineke W H Meijer ⁴, Lioe-Fee de Geus-Oei ^1,², Floris H P van Velden ¹

Editor: Jason Chia-Hsun Hsieh⁵

PMCID: PMC7510999 PMID: 32966313

Abstract

Background

Radiomic features, extracted from positron emission tomography, aim to characterize tumour biology based on tracer intensity, tumour geometry and/or tracer uptake heterogeneity. Currently, radiomic features are derived from static images. However, temporal changes in tracer uptake might reveal new aspects of tumour biology. This study aims to explore additional information of these novel dynamic radiomic features compared to those derived from static or metabolic rate images.

Methods

Thirty-five patients with non-small cell lung carcinoma underwent dynamic [¹⁸F]FDG PET/CT scans. Spatial intensity, shape and texture radiomic features were derived from volumes of interest delineated on static PET and parametric metabolic rate PET. Dynamic grey level cooccurrence matrix (GLCM) and grey level run length matrix (GLRLM) features, assessing the temporal domain unidirectionally, were calculated on eight and sixteen time frames of equal length. Spearman’s rank correlations of parametric and dynamic features with static features were calculated to identify features with potential additional information. Survival analysis was performed for the non-redundant temporal features and a selection of static features using Kaplan-Meier analysis.

Results

Three out of 90 parametric features showed moderate correlations with corresponding static features (ρ≥0.61), all other features showed high correlations (ρ>0.7). Dynamic features are robust independent of frame duration. Five out of 22 dynamic GLCM features showed a negligible to moderate correlation with any static feature, suggesting additional information. All sixteen dynamic GLRLM features showed high correlations with static features, implying redundancy. Log-rank analyses of Kaplan-Meier survival curves for all features dichotomised at the median were insignificant.

Conclusion

This study suggests that, compared to static features, some dynamic GLCM radiomic features show different information, whereas parametric features provide minimal additional information. Future studies should be conducted in larger populations to assess whether there is a clinical benefit of radiomics using the temporal domain over traditional radiomics.

Introduction

In the field of radiomics, researchers aim to find stable and clinically relevant image-derived biomarkers, so-called radiomic features, that provide a non-invasive way of quantifying and monitoring tumour characteristics in clinical practice [1, 2]. For positron emission tomography (PET), radiomic features are investigated that quantify tracer intensity, tumour geometry and/or tracer uptake heterogeneity [3]. Tracer uptake heterogeneity, typically quantified using textural features, describing the spatial distribution of radiotracer uptake of the tumour, might provide information about cellular density, proliferation, angiogenesis, hypoxia, receptor expression, necrosis and fibrosis [4], and it is hypothesized that it, thereby, reflects specific regional variance in tumour characteristics including cancer genetics.

Traditional radiomic features describe heterogeneity along the spatial distribution of radiotracer uptake, but do not take into account tracer uptake heterogeneity over time, while this might contain additional information concerning tumour biology.

Research into these so-called temporal radiomics is limited. There are some studies that apply texture feature analysis on parametric images in magnetic resonance imaging (MRI) [5] and PET [6] using 3D images representing specific pharmacokinetic parameters. However, these studies did not assess time frames as the fourth dimension. To the best of our knowledge, there are no papers describing radiomics in PET using the temporal dimension.

Woods et al. have investigated the use of 4D texture analysis in dynamic contrast-enhanced MRI [7], but interchangeability of spatial and temporal dimensions was assumed. A different approach was found within proteomics, where Hu et al. studied the application of temporal texture features in time series fluorescence microscope images for the analysis of subcellular locations of proteins, since these location patterns indicate the possible function of a protein [8]. They investigated the original thirteen Haralick grey level cooccurrence matrix (GLCM) textural features in the temporal domain. Static GLCM features express combinations of grey levels of neighbouring pixels in the spatial domain [9]; the dynamic approach assesses adjacent voxels in time.

The current study explores whether the temporal domain reflects different aspects of tracer uptake and thereby tumour characteristics compared to static features. Inspired by the approach of Hu et al. [8], we used a different and novel approach, where texture features derived from the temporal domain using dynamic images are developed, i.e. dynamic GLCM features and dynamic grey level run length matrix (GLRLM) features. The aim of this study was to assess the potential additional information content of these dynamic texture features, comparing these to features derived from parametric images and to the more conventional spatial features derived from static images in NSCLC, thereby considering the temporal domain of tracer uptake.

Materials and methods

Patients and clinical follow-up

Dynamic 2-[¹⁸F]fluoro-2-deoxy-D-glucose ([¹⁸F]FDG) positron emission tomography scans combined with X-ray computed tomography (PET/CT) of a previously published prospective cohort [10] were analysed. The study had been reviewed and approved by the Commission on Medical Research Involving Human Subjects Region Arnhem-Nijmegen, the Netherlands. All patients signed an informed consent form. The follow-up data of that previously published prospective cohort were updated. In short, the study included consecutive patients with newly diagnosed or suspected NSCLC, stage IB to stage IIIA, i.e. T2a-4N0-2M0 (TNM 7^th edition), planned for primary resection in the Radboud University Medical Center between 2009 and 2014 [10]. All patients were routinely staged using contrast-enhanced CT of the chest and/or upper abdomen and [¹⁸F]FDG PET/CT with additional histologic staging of the mediastinum or other sites suspicious for cancer when necessary. Only tumours that were considered resectable, with a diameter larger than 30 mm were included to minimize the partial volume effect and be able to quantify heterogeneity [11]. Clinical characteristics of 35 included NSCLC lesions in 34 patients can be found in Table 1.

Table 1. Clinical characteristics of 35 non-small cell lung carcinoma (NSCLC) lesions in 34 patients.

Characteristic	Value
Age (years), median (range)	66 (44–80)
Gender (M/F)	24/11
PET/CT scanner
Biograph Duo	20
Biograph 40 mCT	15
SUV_max (g.mL^-1), median (range)	13.63 (5.62–30.98)
Histology
Squamous cell carcinoma	20
Adenocarcinoma	12
Other	3
Differentiation
Well differentiated	1
Moderately differentiated	13
Poorly differentiated	21
TNM stage (7th edition)
Stage I	8
Stage II	20
Stage III	7
Surgical margins
Free of tumour	32
Not free of tumour	2
Pathologic margins were inconclusive	1
Pleural invasion
Yes	14
No	21
Adjuvant chemotherapy
Yes	17
No	17
Unknown/Lost to follow-up	1
Adjuvant radiation therapy
Yes	4
No	31
Median overall survival (months)	72 (95% CI: 49–95)

Open in a new tab

Patient preparation, data acquisition, image reconstruction and parametrisation

Within 7 days of surgery, patients underwent a dynamic [¹⁸F]FDG PET/CT scan with the primary tumour located centrally in the field of view using either the Biograph Duo (n = 21) or Biograph 40 mCT with TrueV z-axis gantry extension (n = 17) (Siemens Healthineers, Erlangen, Germany). Details on patient preparation, data acquisition and image reconstruction can be found in the original publication [10] and was in accordance with the European Association of Nuclear Medicine (EANM) guidelines for tumour imaging [12]. This resulted in acquisition of 70 time frames representing 60 min of tracer distribution, started at injection of [¹⁸F]FDG.

Static, parametric and dynamic volumes were used in this study and are illustrated in Fig 1. The final time frame (50–60 min p.i.) was used as static [¹⁸F]FDG PET image. The resulting voxel sizes were 2.56×2.56×3.38 and 1.59×1.59×2.03 mm³ for the Biograph Duo PET/CT and Biograph 40 mCT PET/CT, respectively. Parametric glucose metabolic rate (MR_glc) images were computed based on image-derived tissue and blood time-activity concentration curves using Patlak method, with data acquired between 15 and 60 min normalized Patlak-time, using the Patlak slope ([¹⁸F]FDG influx constant, K_i), assuming a lumped constant of 1 and considering the plasma glucose concentration measured prior to [¹⁸F]FDG-injection [10].

Four dimensional (4D) dynamic volumes (x,y,z,t) with time frames of equal acquisition length were created combining the time frames between 10 and 50 min p.i. (16x 75 s; 8x 150 s). Equilibrium between perfusion and uptake of [¹⁸F]FDG is reached after 10–15 min of Patlak time, corresponding to approximately 10–15 min in real time [13]. Different lengths of time frames were created to assess dependence of radiomic features on frame duration, but the options were limited due to the unavailability of the raw data. The time frames were summed, resulting in sixteen 150 s-frames and eight 300 s-frames.

Image analysis

Image acquisition, pre-processing and radiomic feature extraction were performed according to the Image Biomarker Standardisation Initiative (IBSI) guidelines. Details can be found in S1 File.

Volumes of interest

Volumes of interest (VOI) of the tumour in the static and parametric images were drawn independently using a fuzzy locally adaptive Bayesian (FLAB) algorithm [14], excluding [¹⁸F]FDG-avid non-tumour tissue by drawing an oversized container around the tumour and surrounding tissue by a radiation oncologist under supervision of an experienced nuclear medicine physician [10]. The VOIs of the static images were also used as VOIs for all the dynamic images.

Interpolation and discretisation

Interpolation of the image and the VOI was performed to match the voxel sizes of both scanners and to create isotropic voxels, so that image matrices are rotationally invariant [15]. The slice thickness of the Biograph Duo was the largest spatial voxel dimension (3.38 mm). All images were interpolated as recommended by IBSI (trilinearly, grids aligned by centre) to this voxel dimension, i.e. 3.38 × 3.38 × 3.38 mm³ [15] using MATLAB 2017b (Mathworks, Natick, Massachusetts). Dynamic image frames were interpolated before they were combined to a 4D volume.

For the extraction of texture features, grey value discretisation was performed using a fixed bin width, since this leads to more robust features than a fixed number of bins [16]. For standardized uptake value-based images, a bin width of 0.5 g/mL has been described in literature [16]. To the best of our knowledge, optimal bin widths for parametric PET images are not known. Therefore, population-based bin widths were determined according to the Freedman-Diaconis rule [17]:

b i n w i d t h = 2 ∙ I Q R ∙ N^{- 1 / 3}

(1)

with IQR the mean interquartile range and N the mean number of voxels in the VOIs of all included patients. Grey value discretisation of the dynamic images was performed using a fixed bin width for each of the individual time frames, calculated using Eq 1.

Static and parametric features

Radiomic feature extraction of the static and parametric images was performed using PyRadiomics version 2.0 [18] in Python 3.6 (Python Software Foundation, Wilmington, Delaware). For every VOI, 90 features were calculated: intensity (18), shape (13), GLCM (22), GLRLM (16), grey level size zone matrix (GLSZM) (16) and neighbouring grey tone difference matrix (NGDTM) (5). Grey level dependence matrix (GLDM) features were not calculated, since these features are analogues to GLRLM and GLSZM features [15]. Image normalisation and distance weighting were not applied [15]. GLCM matrices were calculated assessing the VOIs in two directions per angle, thus taking into account rotational invariance [15]. GLCM and GLRLM matrices were calculated for thirteen angles (= (3³−1)/2), corresponding to the corresponding direction vectors of the 26 directly neighbouring voxels within a neighbourhood volume at distance 1, and combined to one matrix [15].

Dynamic features

Thirty-eight novel dynamic features were extracted from both the 150 s-frames and the 300 s-frames. Inspired by the approach of Hu et al. [8], novel dynamic texture features were developed that assess changes in voxel values of adjacent voxels in time, only regarding the temporal direction: (x; y; z; t) = (0; 0; 0; 1). By including this temporal direction, the GLCM expresses if voxel values change from one time frame to the next and the GLRLM expresses the frequency of consecutive voxels with the same discretized grey level in the temporal dimension. When changes in tracer uptake over time are limited, the GLCM will only show non-zero values on and near the diagonal and the GLRM will show long runs. Due to causality in the temporal domain GLCMs were only calculated in one direction per angle instead of two [19]. Extraction of 22 dynamic GLCM features and 16 dynamic GLRLM features on these temporal matrices was performed using PyRadiomics version 1.3 [18] in Python 3.6. Dynamic GLSZM and dynamic NGTDM features were not calculated, since these dynamic features would consider spatiotemporal directions instead of solely temporal directions, while spatial and temporal directions are not interchangeable [19]. No differences in implementation between PyRadiomics version 1.3 and 2.0, tested using example data, were found for the radiomic features investigated.

Statistical analysis

Survival data are presented using Kaplan-Meier estimators. Overall survival is defined from date of the [¹⁸F]FDG PET/CT to death of any cause, censoring all patient that were alive at the closeout date (July 30^th 2018).

The dependence of the calculated features on the scanner was assessed using the GlobalTest package in R Statistical Software version 3.3.3 (R Foundation for Statistical Computing, Vienna, Austria) [20]. The association with the used scanner (Biograph Duo or Biograph 40 mCT) was assessed per feature set (static features (90), parametric features (90), dynamic features (38)) and for all features (218) using logistic regression with the features as covariates.

Robustness of dynamic features independent of the length of the time frames was assessed by calculating the Spearman’s rank correlation coefficient (ρ) between the corresponding features from the 150 s- and the 300 s-frames in MATLAB. High correlations (ρ>0.7) indicate robustness of features independent of frame duration. Reproducibility would usually have been assessed using the intraclass correlation coefficient and Bland-Altman analysis, but due to size differences of the GLCM and GLRLM for the different frame durations, systematic scale errors would be expected. In contrast to systematic offset errors, these scale errors cannot be addressed using the conventional statistical methods.

Comparison of parametric and dynamic features with static features was performed using Spearman’s rank correlation coefficient (ρ) in MATLAB. Parametric features with a high correlation (defined as ρ>0.7) with corresponding static features were considered redundant. Only correlations of corresponding features were calculated, since the mathematical definitions of static and parametric features are the same. Dynamic features with a high correlation with any static feature were considered redundant. These redundant features do not contain additional information compared to the original static spatial feature set.

To evaluate clinical relevance, the association of static, parametric and dynamic features with overall survival and histopathological characteristics was assessed in SPSS 23 (IBM Statistics, Chicago, IL). Parametric features and dynamic features that did not show high correlations with static features were assessed. Unsupervised feature selection of the static feature set was performed following the approach of Collarino et al. using redundancy filtering based on Pearson correlation coefficients (r>0.75) and principal component analysis (PCA) [21]. Survival curves were estimated using Kaplan-Meier analysis for the selected features dichotomized at their median and survival curves were compared using log-rank statistics. Differences in radiomic features between different binary histopathological characteristics were statistically compared using the Mann-Whitney U test or independent samples t-test, after testing for (log-)normality.

Results

Extracted radiomic features and clinical data can be found in S1 Data.

Bin widths

The bin width for static and parametric images was 0.55 g/mL and 1.8 × 10⁻⁰⁸ mol/mL/min, respectively. The bin widths for dynamic frames were 0.26, 0.29, 0.31, 0.34, 0.36, 0.36, 0.39, 0.40, 0.42, 0.44, 0.48, 0.49, 0.50, 0.51, 0.53, 0.54 g/mL for the 16 150 s-frames and 0.27, 0.31, 0.35, 0.39, 0.42, 0.48, 0.49, 0.53 g/mL for the 8 300 s-frames.

Differences between scanners

No significant differences between the two scanners (Biograph Duo and Biograph 40 mCT) were found for the static features (p = 0.069), parametric features (p = 0.145), 150 s-frame dynamic features (p = 0.077) and all features together (p = 0.088).

Parametric features

Eighty-seven out of 90 parametric features showed high correlations with corresponding static features, indicating redundancy. Only three out of 90 features did not show high correlations, but showed moderate correlations: short run low grey level emphasis (GLRLM) (ρ = 0.693, P<0.0001), small area emphasis (GLSZM) (ρ = 0.698, P<0.0001) and small area low grey level emphasis (GLSZM) (ρ = 0.610, P = 0.0001).

Dynamic features

Thirty-two out of 38 dynamic features showed very high Spearman’s rank correlations between 150 s- and 300 s-frames. Six GLCM features showed high correlations between frame lengths (ρ≥0.757): Correlation, informational measure of correlation 1 (IMC1), IMC2, inverse difference moment normalized (IDMN), inverse difference normalized (IDN) and inverse variance. Dynamic features were assumed to be robust to a change in frame durations, at least for those investigated in this study. Therefore, the following sections will only show the results for 150 s-frames.

Fig 2 shows the Spearman’s rank correlation matrix of static and dynamic radiomic features derived from the 150 s-time frames. Five dynamic GLCM features show a negligible to moderate correlation (ρ<0.7) with any static feature, indicating potential additional information to existing features. The maximum correlation between any static feature and the dynamic GLCM features IMC1 and correlation were negligible to low (ρ = 0.181 and = 0.483, respectively). Three features IDN, IDMN and IMC2 maximally showed a moderate correlation with any static features (ρ = 0.533, ρ = 0.542 and ρ = 0.549, respectively). All other dynamic GLCM and dynamic GLRLM showed a high correlation (ρ>0.7) with at least one static feature.

Clinical relevance

PCA was inconclusive, since the number of subjects was too low compared to the number of features in order to obtain significant results. Therefore, the remaining five dynamic and three parametric features were compared to three traditional quantitative PET features: the maximum standardized uptake value (SUV_max), the metabolically active tumour volume (MTV) and the total lesion glycolysis (TLG). The estimated Kaplan-Meier survival curves for overall survival (OS) were not significantly different for these selected radiomic features dichotomized at their median. Kaplan-Meier survival curves can be found in S1 Fig and associations of radiomic features with histopathological characteristics can be found in S1 Table. The static SUV_max, the parametric short run low grey level emphasis (GLRLM) and the parametric small area low grey level emphasis (GLRLM) showed significant differences between adenocarcinomas and squamous cell carcinomas. Static MTV and TLG showed borderline significant differences in pleural invasion.

Discussion

In this study, additional information and redundancy of novel radiomic features describing the temporal domain compared to traditional radiomic features derived from static images were investigated in a dataset of 35 lesions in 34 patients with stage IB to IIIA NSCLC who underwent a dynamic [¹⁸F]FDG PET/CT scan.

To the best of our knowledge, there are no papers describing radiomic features assessing tracer uptake heterogeneity over time in PET imaging. Woods et al. [7] investigated the addition of the temporal domain in radiomics. However, interchangeability of spatial and temporal dimensions was assumed, which might be questionable, since causality is a condition in the temporal dimension, while it is not in the spatial dimensions [19]. Therefore, inspired by the approach of Hu et al. in proteomics [8], where dynamic GLCM features were designed to assess combinations of grey levels of subsequent voxels in time, we developed novel dynamic GLCM and GLRLM features.

Five out of 22 dynamic GLCM features did not show redundancy in comparison to any static feature, suggesting potential additional information compared to static features. The dynamic GLRLM features showed high correlation with static features, demonstrating that these features likely do not include additional information concerning tracer uptake compared to the traditional radiomic feature set. Dynamic features were robust independent of frame length. Features from parametric images showed, except for three out of 90 features, high correlations with their static equivalents, demonstrating minimal additional information. These findings show that some dynamic GLCM features might potentially express additional information about tracer uptake, suggesting potential insights in tumour biology. Combined with the static feature set, they might provide extensive and in-depth knowledge in biological processes.

Unfortunately we could not show the clinical relevance using survival analysis in our small dataset. Parametric and dynamic as well as traditional quantitative PET features dichotomized at the median, did not show significant differences in Kaplan-Meier curves between the high and the low group. Intriguingly, even the quantitative feature SUV_max did not show a difference, while high values of SUV_max predicted a higher risk of death in patients with NSCLC [22], also in patients with the same stage as our cohort, treated with a surgical resection [23]. Two parametric features and the SUV_max did show significant differences between adenocarcinomas and squamous cell carcinomas, but these features cannot be used for patient stratification, since the ranges of the features overlap. Since clinical added value could not be demonstrated, it is unknown whether the different information that the non-redundant features contained, has an added value interpreting tumour biology or consists of merely noise. Consequently, future dynamic PET studies with a large number of patients are warranted to investigate whether these dynamic radiomic features show any correlations with clinical outcome measures and are useful for predictive or prognostic purposes, especially, considering the development of more sophisticated machine learning and deep learning algorithms, even requiring larger numbers of patients [24].

In extension to Tixier et al. [6], correlations between radiomic features derived from static and parametric images were assessed side-by-side for this larger feature set. Correlations found by Tixier et al. cannot be compared one-to-one with our results, since different settings for radiomic feature extraction were used, e.g. a fixed number of bins versus a fixed bin width. Nevertheless, results of both studies show the same trend. All eight features assessed by Tixier et al. show high correlations between static and parametric radiomics. In our study, except for three features that were not assessed by Tixier et al., all other parametric features showed a high correlation with their static equivalents. These three features showed moderate correlations (ρ>0.610), only suggesting a minimal amount of additional information. It should be noted, though, that the study by Tixier et al. only included the first 20 patients of the cohort of the current study performed on a single PET-scanner (Biograph Duo).

Bin widths were calculated using the Freedman-Diaconis rule [17], since, to the best of our knowledge, bin widths for parametric images and dynamic frames cannot be found in literature. Another approach could be the calculation of the regression coefficient of the static SUV_mean and mean glucose metabolic rate or SUV_mean of a dynamic frame for all patients. This regression coefficient and a static bin width of 0.5 g/mL could be used to calculate the parametric and dynamic frame bin widths. In our population, the regression coefficient of the SUV_mean and mean glucose metabolic rate was 0.0296 (strong linear correlation, r = 0.93). The static bin width of 0.55 g/mL, as calculated using the Freedman-Diaconis rule, would result in a bin width of 0.016 μmol/mL/min using the regression coefficient, which is very similar to the bin width of 0.018 μmol/mL/min calculated using the Freedman-Diaconis rule. The bin widths for the dynamic frames calculated using both methods were also similar.

Despite this paper lacking clinical relevance, in our opinion, the extraction of dynamic radiomic features might complement the static feature set and thereby advance the field by providing insight in different aspects of tumour biology. In addition, the field suffers from a publication bias with only 6% of radiomic papers describing negative results, as stretched by Buvat et al. [25]. Clinical validation and implementation in decision support of dynamic radiomic features might be difficult, as it requires relatively large datasets and dynamic PET acquisition is not standard-of-care due to high costs and invasive nature. This contradicts one of the main goals of radiomics of extracting biomarkers from standard-of-care medical images, lowering the need for additional or more invasive diagnostic methods. However, the recent introduction of total body PET scanners might facilitate acquisition of dynamic scans in clinical practice [26]. Also, extraction of texture features from dynamic images might be interesting in other modalities like dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) and DCE-CT.

Conclusion

In dynamic [¹⁸F]FDG PET/CT scans in patients with non-small cell lung carcinoma, certain dynamic GLCM radiomic features show different information than traditional radiomic features. These novel dynamic features are robust to an alternation in the frame duration. Features from parametric images only demonstrated minimal additional information. Future studies should assess whether there is a clinical benefit of radiomic features from dynamic images compared to traditional features derived from static images.

Supporting information

S1 File. Image biomarker standardisation initiative reporting guidelines.

(DOCX)

Click here for additional data file.^{(51.1KB, docx)}

S1 Data. Data file radiomic features and clinical data.

(XLSX)

Click here for additional data file.^{(146KB, xlsx)}

S1 Fig. Estimated Kaplan-Meier overall survival curves for selected features dichotomized at the median, compared using log-rank statistics.

(DOCX)

Click here for additional data file.^{(1.2MB, docx)}

S1 Table. Association between histopathological characteristics and radiomic features calculated using the Mann-Whitney U test or independent samples t-test, after testing for (log-)normality.

(DOCX)

Click here for additional data file.^{(18.8KB, docx)}

Acknowledgments

The authors want to thank Nicolle Peters for help with patient inclusion and Peter Kok and his team of PET technologists for support during acquisition of the original study.

List of abbreviations

4D: four dimensional
CT: computed tomography
DCE: dynamic contrast enhanced
EANM: European Association of Nuclear Medicine
[¹⁸F]FDG: 2-[¹⁸F]fluoro-2-deoxy-D-glucose
FLAB: fuzzy locally adaptive Bayesian
GLCM: grey level cooccurrence matrix
GLDM: grey level dependence matrix
GLRLM: grey level run length matrix
GLSZM: grey level size zone matrix
IBSI: Image Biomarker Standardisation Initiative
IDMN: inverse difference moment normalized
IDN: inverse difference normalized
IMC: informational measure of correlation
MR_glc: parametric glucose metabolic rate
MRI: magnetic resonance imaging
OS: overall survival
NGTDM: neighbouring grey tone difference matrix
NSCLC: non-small cell lung carcinoma
PET: positron emission tomography
VOI: volume of interest

Data Availability

All relevant data are within the manuscript and its Supporting Information files.

Funding Statement

Dennis Vriens was supported in part by the Netherlands Organisation for Health Research and Development (ZonMw) stipends for Qualified Doctor Training to become a Clinical Researcher (AGIKO) (project no. 92003552) for design and data collection of the original clinical study. The costs of the additional dynamic PET scans were covered by the Department of Radiology and Nuclear Medicine, Radboud University Medical Center, Nijmegen. There was no additional funding received for this study.

References

1.Parmar C, Grossmann P, Bussink J, Lambin P, Aerts HJWL. Machine Learning methods for Quantitative Radiomic Biomarkers. Sci Rep. 2015;5:13087 10.1038/srep13087 [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Limkin EJ, Sun R, Dercle L, Zacharaki EI, Robert C, Reuze S, et al. Promises and challenges for the implementation of computational medical imaging (radiomics) in oncology. Ann Oncol. 2017;28(6):1191–206. 10.1093/annonc/mdx034 [DOI] [PubMed] [Google Scholar]
3.Hatt M, Tixier F, Pierce L, Kinahan PE, Le Rest CC, Visvikis D. Characterization of PET/CT images using texture analysis: the past, the present… any future? Eur J Nucl Med Mol Imaging. 2017;44(1):151–65. 10.1007/s00259-016-3427-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Junttila MR, de Sauvage FJ. Influence of tumour micro-environment heterogeneity on therapeutic response. Nature. 2013;501:346 10.1038/nature12626 [DOI] [PubMed] [Google Scholar]
5.Kassner A, Thornhill RE. Texture Analysis: A Review of Neurologic MR Imaging Applications. AJNR Am J Neuroradiol. 2010;31(5):809–16. 10.3174/ajnr.A2061 [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Tixier F, Vriens D, Cheze-Le Rest C, Hatt M, Disselhorst JA, Oyen WJ, et al. Comparison of Tumor Uptake Heterogeneity Characterization Between Static and Parametric 18F-FDG PET Images in Non-Small Cell Lung Cancer. J Nucl Med. 2016;57(7):1033–9. 10.2967/jnumed.115.166918 [DOI] [PubMed] [Google Scholar]
7.Woods BJ, Clymer BD, Kurc T, Heverhagen JT, Stevens R, Orsdemir A, et al. Malignant-lesion segmentation using 4D co-occurrence texture analysis applied to dynamic contrast-enhanced magnetic resonance breast image data. J Magn Reson Imaging. 2007;25(3):495–501. 10.1002/jmri.20837 [DOI] [PubMed] [Google Scholar]
8.Hu Y, Carmona J, Murphy RF, editors. Application of temporal texture features to automated analysis of protein subcellular locations in time series fluorescence microscope images 2006 IEEE International Symposium on Biomedical Imaging; 2006: IEEE. [Google Scholar]
9.Haralick RM, Shanmugam K, Dinstein I. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics. 1973;SMC-3(6):610–21. [Google Scholar]
10.Meijer TWH, de Geus-Oei LF, Visser EP, Oyen WJG, Looijen-Salamon MG, Visvikis D, et al. Tumor Delineation and Quantitative Assessment of Glucose Metabolic Rate within Histologic Subtypes of Non-Small Cell Lung Cancer by Using Dynamic (18)F Fluorodeoxyglucose PET. Radiology. 2017;283(2):547–59. 10.1148/radiol.2016160329 [DOI] [PubMed] [Google Scholar]
11.Hatt M, Majdoub M, Vallieres M, Tixier F, Le Rest CC, Groheux D, et al. 18F-FDG PET uptake characterization through texture analysis: investigating the complementary nature of heterogeneity and functional tumor volume in a multi-cancer site patient cohort. J Nucl Med. 2015;56(1):38–44. 10.2967/jnumed.114.144055 [DOI] [PubMed] [Google Scholar]
12.Boellaard R, O'Doherty MJ, Weber WA, Mottaghy FM, Lonsdale MN, Stroobants SG, et al. FDG PET and PET/CT: EANM procedure guidelines for tumour PET imaging: version 1.0. Eur J Nucl Med Mol Imaging. 2010;37(1):181–200. 10.1007/s00259-009-1297-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Vriens D, Visser EP, de Geus-Oei LF, Oyen WJ. Methodological considerations in quantification of oncological FDG PET studies. Eur J Nucl Med Mol Imaging. 2010;37(7):1408–25. 10.1007/s00259-009-1306-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Hatt M, Cheze le Rest C, Turzo A, Roux C, Visvikis D. A fuzzy locally adaptive Bayesian segmentation approach for volume determination in PET. IEEE Trans Med Imaging. 2009;28(6):881–93. 10.1109/TMI.2008.2012036 [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Zwanenburg A, Leger S, Vallieres M, Lock S. Image biomarker standardisation initiative—feature definitions v11. CoRR. 2019;1612.07003. [Google Scholar]
16.Leijenaar RT, Nalbantov G, Carvalho S, van Elmpt WJ, Troost EG, Boellaard R, et al. The effect of SUV discretization in quantitative FDG-PET Radiomics: the need for standardized methodology in tumor texture analysis. Sci Rep. 2015;5:11075 10.1038/srep11075 [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Yip SS, Aerts HJ. Applications and limitations of radiomics. Phys Med Biol. 2016;61(13):R150–66. 10.1088/0031-9155/61/13/R150 [DOI] [PMC free article] [PubMed] [Google Scholar]
18.van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. 2017;77(21):e104–e7. 10.1158/0008-5472.CAN-17-0339 [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Bianco-Martinez E, Baptista MS. Space-time nature of causality. Chaos: An Interdisciplinary Journal of Nonlinear Science. 2018;28(7). [DOI] [PubMed] [Google Scholar]
20.Goeman JJ, van de Geer SA, de Kort F, van Houwelingen HC. A global test for groups of genes: testing association with a clinical outcome. Bioinformatics. 2004;20(1):93–9. 10.1093/bioinformatics/btg382 [DOI] [PubMed] [Google Scholar]
21.Collarino A, Garganese G, Fragomeni SM, Pereira Arias-Bouda LM, Ieria FP, Boellaard R, et al. Radiomics in vulvar cancer: first clinical experience using (18)F-FDG PET/CT images. J Nucl Med. 2019;60(2):199–206. [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Liu J, Dong M, Sun X, Li W, Xing L, Yu J. Prognostic Value of 18F-FDG PET/CT in Surgical Non-Small Cell Lung Cancer: A Meta-Analysis. PLoS One. 2016;11(1):e0146195 10.1371/journal.pone.0146195 [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Ohtaka K, Hida Y, Kaga K, Okamoto S, Shiga T, Tamaki N, et al. Outcome Analysis of 18F-Fluorodeoxyglucose Positron–Emission Tomography in Patients with Lung Cancer After Partial Volume Correction. 2013;33(11):5193–8. [PubMed] [Google Scholar]
24.Razzak MI, Naz S, Zaib A. Deep Learning for Medical Image Processing: Overview, Challenges and the Future. In: Dey N, Ashour AS, Borra S, editors. Classification in BioApps: Automation of Decision Making. Cham: Springer International Publishing; 2018. p. 323–50. [Google Scholar]
25.Buvat I ORLHAC F. The dark side of radiomics: on the paramount importance of publishing negative results. J Nucl Med. 2019. [DOI] [PubMed] [Google Scholar]
26.Cherry SR, Jones T, Karp JS, Qi J, Moses WW, Badawi RD. Total-Body PET: Maximizing Sensitivity to Create New Opportunities for Clinical Research and Patient Care. J Nucl Med. 2018;59(1):3–12. 10.2967/jnumed.116.184028 [DOI] [PMC free article] [PubMed] [Google Scholar]

PLoS One. doi: 10.1371/journal.pone.0239438.r001

Decision Letter 0

Jason Chia-Hsun Hsieh

5 Jun 2020

PONE-D-20-05431

Adding the temporal domain to PET radiomic features

PLOS ONE

Dear Dr. Noortman,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

The study is interesting and might be important in this field. Please kindly address the issues from the reviewers.

Please submit your revised manuscript by Jul 20 2020 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols

We look forward to receiving your revised manuscript.

Kind regards,

Jason Chia-Hsun Hsieh, M.D. Ph.D

Academic Editor

PLOS ONE

Additional Editor Comments:

The study is interesting and might be important in this field.

Journal requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. In the ethics statement in the manuscript and in the online submission form, please provide additional information about the patient records used in your retrospective study, including:

a) whether all data were fully anonymized before you accessed them and

b) the date range (month and year) during which patients' medical records were accessed.

3. Please confirm in your methods section and ethics statement that the 'Commission on Medical Research Involving Human Subjects Region Arnhem-Nijmegen' consists of a committee of experts that reviewed and approved your study.

In addition, please clarify whether the present retrospective study was also granted ethical approval, in addition to the original study, and whether the previous prospective cohort were recruited by the same authors.

4. Thank you for stating in your Funding Statement:

'Dennis Vriens was supported in part by the Netherlands Organisation for Health Research and Development (ZonMw) stipends for Qualified Doctor Training to become a Clinical Researcher (AGIKO) (project no. 92003552) for design and data collection of the original clinical study.'

a. Please provide an amended statement that declares *all* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now

Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement.

b. Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf.

5. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: http://journals.plos.org/plosone/s/supporting-information

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

Reviewer #4: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: Yes

Reviewer #4: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: No

Reviewer #4: No

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

Reviewer #4: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: The authors presented the possibility of adding PET radiomics features by additional information of temporal dynamic PET images. Currently, static information was used in the texture analysis for the radiomics analysis. However, kinetic information from the PET study may have significant data as a phenotyping information.

The authors describe the possibility of the use in PET radiomics using dynamic and kinetic data on PET.

They included only a 35-patients with NSCLC and dynamic FDG PET/CT, many kinds of additional parameters may have a additional information compared to the static PET images.

This is an preliminary report, but well-designed, prospective study with a meaningful suggestion for the use of total body PET with real dynamic PET study. Thus, this paper is very helpful for the development of imaging biomarker as well as PET radiomics phenotyping study.

Reviewer #2: This manuscript focuses on the development of a new method to evaluate the tumor heterogeneity based on dynamic 18F-FDG PET/CT images. Authors showed that new “dynamic” features can be extracted with a moderate correlation with “static” features.

Even if this new approach is promising, authors failed to demonstrate any added value. This may be because the cohort is too small or because the features do not carry information related to the patient survival. Worse, perhaps these new features do not reflect relevant biological information. Did the authors test the link between the “dynamic” features and the tumor histology or the differentiation for instance?

For the survival analysis, the authors do not seem to have taken into account the pleural invasion or the treatment (chemotherapy and/or radiation therapy), event though it is known that this strongly influences the prognosis.

This study may be too preliminary and requires further investigations to demonstrate a clinical benefit.

Reviewer #3: The paper addressed radiomics analysis on dynamic PET studies and include the temporal domain (either by using dynamic frames or using patlak Ki images). The paper is of interest as it generates some hypotheses and new ideas on the use of dynamic information for radiomics studies. As there is not much clinical benefit I would recommend to also describe the difference in RF values when analysed on Ki and SUV images (matching the number of bins used, see below)

Main comments:

VOIs – it seems that previously defined vois were reused or were these newly drawn?. Any differences in voi when defined on static and parametric images?. Did the authors redraw voi in the dynamic frames? If not, which one was used for dynamic analysis?

Bin width – I understand that for parametric images and for dynamic frames you cannot use the SUV=0.5 bin width, but you can estimate the slope between SUV and KI and then estimate how to convert SUV=0.5 bins into Ki bins. I recommend to do such an analysis to see if RF values become more comparable when you try to match the bin width (taking into account the different units of SUV and Ki).

Likewise, you can derive how many bins you got per static SUV image and use that number to process the dynamic frames or parametric images (for that patient).

Apart from assessing spearman correlations, it would be nice to demonstrate the difference in RF values, eg by using a distance metric or else, but it will likely require the above suggested bin width adaptions as well.

You have mainly stage 2 subjects. There is no difference in KM plots. Maybe add some case control evaluations by looking at the feature values for stage 1 versus stage 3 and see if there are any features that are significantly different between these 2 more extreme cases…I realize it is only about 7 subjects per group (in this case), but it would give an hypothesis if some features might have prognostic value?

Likewise, you can take the 25% short survivors and 25% long survivors and see if there is any difference between these groups.

Minor comments:

Abstract,results: 3 out of 90 show moderate correlation, so the others were highly correlated. please state so.

Two scanners are used. Was there any cross-calibration between the systems. They do not find significant differences (yet a trend, so there must be some difference there?).

Patlak images – which software was used?

2 software packages were used for RF. Any chance of different implementation issues? Did the authors compare results from the 2 packages.

Reviewer #4: In the manuscript

„Adding the temporal domain to PET radiomic features”,

the authors analyze the effect of introducing the temporal domain in the assessment of radiomics features in PET data. This is a new and innovative idea in radiomics analysis of PET data and of high interest for the community. The manuscript is very well written and good to understand. I just have some minor issues that should be changes:

- How were the blood time-activity concentration curves estimated, by blood-sampling or image-based; if so, was the left ventricle used?

- The resolution of figure 4 is way to low, it is hardly possible to see anything. Pleas also include the risk tables below the Kaplan-Meier curves.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

Reviewer #4: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2020 Sep 23;15(9):e0239438. doi: 10.1371/journal.pone.0239438.r002

Author response to Decision Letter 0

23 Jul 2020

Article ID: PONE-D-20-05431

Title: Adding the temporal domain to PET radiomic features

We would like to thank the reviewers for their helpful comments which have improved the quality of the paper. The specific comments of the reviewers are in blue/italic below together with their responses. Unfortunately, we did not manage to get the text in blue/italic in this editor, but this version can be found in the attached files. We have changed the manuscript accordingly (changes are marked in red).

Response to Journal requirements:

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

We apologize for missing these files when submitting the manuscript. The manuscript is now updated according to the style requirements.

2. In the ethics statement in the manuscript and in the online submission form, please provide additional information about the patient records used in your retrospective study, including:

a) whether all data were fully anonymized before you accessed them and

b) the date range (month and year) during which patients' medical records were accessed.

After informed consent, relevant patient data were included in the case report forms by the treating physicians. Follow-up data were updated in July 2018. Clinical data were pseudonymized before radiomic analysis. Imaging data (DICOM files) were left unchanged, as advanced quantitative analysis was part of the original study for which the patients provided written informed consent and as DICOM anonymization might remove DICOM tags that are crucial for absolute [18F]FDG quantification. After data extraction all further analyses were performed pseudonymized. Imaging data were accessed between January and October 2018. Pseudonymized clinical data and radiomic features were assessed between January 2018 and February 2020.

In the present study, we performed an additional analysis of this previously published cohort. For the current study ethical approval was granted within the approval of the original prospective study. The application included quantitative image analysis, as implemented in this study. The prospective cohort was recruited by the same authors (DV, JB, TM and LF).

4. Thank you for stating in your Funding Statement:

Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement.

b. Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf.

Thank you, the funding statement was changed.

Thank you for pointing this out, the supporting information files were added as captions at the end of the manuscript, in line with the Supporting Information guidelines.

Response to comments to the author:

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

Reviewer #4: Yes

________________________________________

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: Yes

Reviewer #4: Yes

________________________________________

3. Have the authors made all data underlying the findings in their manuscript fully available?

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: No

Reviewer #4: No

________________________________________

4. Is the manuscript presented in an intelligible fashion and written in standard English?

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

Reviewer #4: Yes

________________________________________

5. Review Comments to the Author

The authors describe the possibility of the use in PET radiomics using dynamic and kinetic data on PET.

They included only a 35-patients with NSCLC and dynamic FDG PET/CT, many kinds of additional parameters may have a additional information compared to the static PET images.

We would like to thank the reviewer for the compliments and endorsement of the potential benefit of these novel dynamic radiomics features in future total body PET studies.

We would like to thank the reviewer for this suggestion. Unfortunately, homogeneous and large cohorts of patients who underwent dynamic PET scans are scarce. We investigated the association of the parametric, dynamic and traditional quantitative PET features with clinical parameters. We added supporting file 3 to the manuscript, where these findings are presented. Table 1 of S3 presents the association with clinical parameters. Two parametric features and the SUVmax showed significant differences in mean between adenocarcinoma and squamous cell carcinoma, but the ranges of the values are overlapping, indicating that these parameters cannot be used to discriminate between both histopathological subtypes. These findings were added to the manuscript (lines: 236-238, 287-290, 331-333).

The reviewer is correct. We did not take into account these factors, because we only performed Kaplan-Meier analysis. In table 1 we added the univariate and multivariate Cox regression analysis of the clinical characteristics. In our population, only one variable, the age of diagnosis, is significantly associated with survival.

Table 1: Univariate and multivariate Cox regression analysis of clinical characteristics, traditional quantitative PET features and selected radiomic features for overall survival (OS), disease-free survival (DFS) and disease-specific survival (DSS). Characteristics and features with a p-value < 0.20 in univariate analysis, were selected for forward and backward multivariate analysis based on the likelihood ratio. Pleural invasion and adjuvant chemotherapy were removed from the model in both forward and backward multivariate Cox regression.

Hazard ratio (95% confidence interval) p-value

Univariate Cox regression analysis

Gender (male/female) 1.937 (0.545 - 6.889) 0.298

Age of diagnosis (years) 1.086 (1.014 - 1.164) 0.015

Stadium

IB 1.000 0.697

IIA 1.514 (0.252 - 9.100)

IIB 1.081 (0.198 - 5.922)

IIIA 2.233 (0.449 - 11.112)

Pleural invasion 0.486 (1.175 - 1.349) 0.157

Negative resection margin 0.594 (0.074 - 4.756) 0.620

Adjuvant chemotherapy 2.390 (0.796 - 7.181) 0.109

Post-operative radiotherapy 0.849 (0.189 - 3.809) 0.830

Traditional quantitative PET features

SUVmax (g/mL) 1.029 (0.957 - 1.106) 0.432

MTV (mL) 1.000 (1.000 - 1.000) 0.606

TLG (g) 1.000 (1.000 - 1.000) 0.611

Parametric features

GLRLM SRLGLE 5.645 (0.000 - 94206.982) 0.727

GLSZM SAE 17.494 (0.099 - 3083.885) 0.279

GLSZM SALGLE 1.615 (0.016 - 160.357) 0.838

Dynamic GLCM

Correlation 0.253 (0.000 - 1287.290) 0.253

IMC1 18.304 (0.002 - 149105.596) 0.526

IMC2 1.925 (0.000 - 182838.748) 0.911

IDMN 0.000 (0.000 - 4.23E+73) 0.714

IDN 0.000 (0.000 - 2.339E+17) 0.663

Multivariate Cox regression analysis

Iterative forward selection

Age of diagnosis (years) 1.083 (1.009 - 1.163) 0.027

Pleural invasion

Adjuvant chemotherapy

Iterative backward selection

Age of diagnosis (years) 1.083 (1.009 - 1.163) 0.027

Pleural invasion

Adjuvant chemotherapy

This study may be too preliminary and requires further investigations to demonstrate a clinical benefit.

We would like to thank the reviewer for the helpful suggestions and hope that by adding these additional analyses to the manuscript we have improved its quality. We agree that the study is too preliminary to show clinical benefit and that this study only shows that some dynamic GLCM radiomic features show different information. We further agree that future studies with a larger cohort should be conducted to show the additional clinical benefit of these dynamic features. This is acknowledged in the discussion.

We would like to thank the reviewer for the helpful recommendations. We hope to have sufficiently addressed all of the reviewers concerns below.

Main comments:

The reviewer is correct, we reused the VOIs that were previously defined in the original study. These were defined on the static images and parametric images separately. For the current analysis, we copied them unchanged to the static and parametric images. To avoid that the delineation would change between different frames and thereby impact feature extraction, we reused the static VOIs as VOIs for the dynamic frames. So these VOIs were also not redrawn. This can be found in lines 152-156 of the manuscript.

Likewise, you can derive how many bins you got per static SUV image and use that number to process the dynamic frames or parametric images (for that patient).

Thank you for the interesting suggestion, we did not consider this. The graph in figure 1 (left) shows how the SUVmean translates to the mean glucose metabolic rate of the parametric images for all patients in our study. We used the slope of this graph to calculate the parametric bin width, which would be 0.0296*0.5 = 0.016 µmol/mL/min. We used the same approach for the dynamic frames, as an example, figure 1 (right) shows how the static SUVmean translates to the SUVmean in frame 62, which would result in a bin width of 0.7038*0.5 = 0.35 g/mL. We did this for all dynamic frames. It turns out that these bin widths are very similar to the bin widths calculated with the Freedman-Diaconis rule, especially, when we calculate the bin widths using the slope and a static bin width of 0.55 g/mL, as used in our study. Table 2 shows the calculated slopes and bin widths and the bin widths calculated with the Freedman-Diaconis rule, for all used images. Since these values are quite similar, we did not change the bin widths, but we did mention this in our manuscript (lines: 354-363).

Figure 1. Right: Calculation of the slope of the static SUVmean and the mean glucose metabolic rate for all subjects. Left: Calculation of the slope of the static SUVmean and the SUVmean of dynamic frame 62. The slopes can be used to translate the bin width of the static images to parametric and dynamic bin widths, respectively.

Table 2: Calculated slopes and slope bin widths and bin widths calculated using the Freedman-Diaconis rule.

Image Static MRglu F46 F48 F50 F52 F54 F56 F58 F60

Slope 0.0296 0.4317 0.4666 0.5017 0.5442 0.5723 0.606 0.6434 0.6755

Bin width slope 0.55 0.016 0.24 0.26 0.28 0.30 0.32 0.33 0.35 0.37

Bin width Freedman-Diaconis 0.55 0.018 0.26 0.29 0.31 0.34 0.36 0.36 0.37 0.4

Image F62 F63 F64 F65 F66 F67 F68 F69

Slope 0.7038 0.7368 0.7542 0.7794 0.8025 0.8189 0.856 0.8814

Bin width slope 0.39 0.40 0.42 0.43 0.44 0.45 0.47 0.49

Bin width Freedman-Diaconis 0.42 0.44 0.48 0.49 0.5 0.51 0.53 0.54

In the past we have investigated comparing features using distance metrics, but it is difficult to interpret these results for the different radiomic features, since mathematical definitions vary largely from feature to feature.

Likewise, you can take the 25% short survivors and 25% long survivors and see if there is any difference between these groups.

We would like to thank the reviewer for this suggestion. We added supporting file 3 to the manuscript. Table 1 of S3 presents the association with clinical parameters. Unfortunately, no significant differences were found.

Minor comments:

Abstract,results: 3 out of 90 show moderate correlation, so the others were highly correlated. please state so.

We added this for clarity (line 43).

Two scanners are used. Was there any cross-calibration between the systems. They do not find significant differences (yet a trend, so there must be some difference there?).

The scanners were not used at the same time, so cross-calibration between the scanners has not been performed. However, both scanners were EARL accredited and were cross-calibrated with the dose calibrator. The same dose calibrator was used throughout the whole study.

Patlak images – which software was used?

The Patlak analysis was performed in Inveon Research Workplace (Siemens Healthineers, Erlangen, Germany). This was added to Appendix S1, which contains more information on image acquisition and reconstructions.

2 software packages were used for RF. Any chance of different implementation issues? Did the authors compare results from the 2 packages.

The main difference between PyRadiomics 1.3 and 2.0 is the calculation of the matrices from which the features are extracted. In PyRadiomics 1.3, these matrices are calculated in Python, while in PyRadiomics 2.0, the matrices are calculated in C. We adjusted PyRadiomics 1.3 for the extraction of the dynamic features. For comparison, table 3 presents radiomic features calculated for the PyRadiomics example data, for version 1.3 and 2.0. No differences were found between both implementations. This was added in line 200 of the manuscript.

Table 3. Feature values for PyRadiomics version 1.3 and 2.0 and differences between implementations.

Feature Version 1.3 Version 2.0 Difference

Image lung1_image.nrrd lung1_image.nrrd

Mask lung1_label.nrrd lung1_label.nrrd

general_info_BoundingBox (206, 347, 32, 24, 26, 3) (206, 347, 32, 24, 26, 3)

general_info_EnabledImageTypes {'Original': {}} {'Original': {}}

general_info_GeneralSettings {'minimumROIDimensions': 1, 'minimumROISize': None, 'normalize': False, 'normalizeScale': 1, 'removeOutliers': None, 'resampledPixelSpacing': None, 'interpolator': 'sitkBSpline', 'preCrop': False, 'padDistance': 5, 'distances': [1], 'force2D': False, 'force2Ddimension': 0, 'resegmentRange': None, 'label': 1, 'additionalInfo': True, 'voxelBased': False} {'minimumROIDimensions': 1, 'minimumROISize': None, 'normalize': False, 'normalizeScale': 1, 'removeOutliers': None, 'resampledPixelSpacing': None, 'interpolator': 'sitkBSpline', 'preCrop': False, 'padDistance': 5, 'distances':

general_info_ImageHash 34dca4200809a5e76c702d6b9503d958093057a3 34dca4200809a5e76c702d6b9503d958093057a3

general_info_ImageSpacing (0.5703125, 0.5703125, 5.0) (0.5703125, 0.5703125, 5.0)

general_info_MaskHash 054d887740012177bd1f9031ddac2b67170af0f3 054d887740012177bd1f9031ddac2b67170af0f3

general_info_NumpyVersion 1.19.0 1.19.0

general_info_PyWaveletVersion 1.1.1 1.1.1

general_info_SimpleITKVersion 1.2.4 1.2.4

general_info_Version 1.3.0 2.0.0

general_info_VolumeNum 1 1

general_info_VoxelNum 837 837

original_shape_Elongation 0.718791031 0.718791031 0

original_shape_Flatness 0.514335768 0.514335768 0

original_shape_LeastAxis 8.936318224 8.936318224 0

original_shape_MajorAxis 17.37448332 17.37448332 0

original_shape_Maximum2DDiameterColumn 16.04444054 16.04444054 0

original_shape_Maximum2DDiameterRow 13.53756348 13.53756348 0

original_shape_Maximum2DDiameterSlice 15.97893091 15.97893091 0

original_shape_Maximum3DDiameter 18.18259471 18.18259471 0

original_shape_MinorAxis 12.48862278 12.48862278 0

original_shape_Sphericity 0.75931875 0.75931875 0

original_shape_SurfaceArea 782.241458 782.241458 0

original_shape_SurfaceVolumeRatio 0.574671403 0.574671403 0

original_shape_Volume 1361.197815 1361.197815 0

original_firstorder_10Percentile -245.4 -245.4 0

original_firstorder_90Percentile 71 71 0

original_firstorder_Energy 16291991 16291991 0

original_firstorder_Entropy 4.020834927 4.020834927 0

original_firstorder_InterquartileRange 198 198 0

original_firstorder_Kurtosis 2.695927096 2.695927096 0

original_firstorder_Maximum 106 106 0

original_firstorder_MeanAbsoluteDeviation 105.0944475 105.0944475 0

original_firstorder_Mean -63.9080048 -63.9080048 0

original_firstorder_Median -31 -31 0

original_firstorder_Minimum -506 -506 0

original_firstorder_Range 612 612 0

original_firstorder_RobustMeanAbsoluteDeviation 81.58090535 81.58090535 0

original_firstorder_RootMeanSquared 139.5161078 139.5161078 0

original_firstorder_Skewness -0.73366595 -0.73366595 0

original_firstorder_TotalEnergy 26495367.44 26495367.44 0

original_firstorder_Uniformity 0.074426645 0.074426645 0

original_firstorder_Variance 15380.51125 15380.51125 0

original_glcm_Autocorrelation 411.4164748 411.4164748 0

original_glcm_ClusterProminence 9732.694396 9732.694396 0

original_glcm_ClusterShade -345.713367 -345.713367 0

original_glcm_ClusterTendency 58.74756668 58.74756668 0

original_glcm_Contrast 20.7134493 20.7134493 0

original_glcm_Correlation 0.470613617 0.470613617 0

original_glcm_DifferenceAverage 3.216603092 3.216603092 0

original_glcm_DifferenceEntropy 3.187524502 3.187524502 0

original_glcm_DifferenceVariance 9.381995813 9.381995813 0

original_glcm_Id 0.417361964 0.417361964 0

original_glcm_Idm 0.344350177 0.344350177 0

original_glcm_Idmn 0.972582394 0.972582394 0

original_glcm_Idn 0.899630707 0.899630707 0

original_glcm_Imc1 -0.17331187 -0.17331187 0

original_glcm_Imc2 0.818766382 0.818766382 0

original_glcm_InverseVariance 0.278697167 0.278697167 0

original_glcm_JointAverage 20.04512484 20.04512484 0

original_glcm_JointEnergy 0.017918271 0.017918271 0

original_glcm_JointEntropy 6.932828996 6.932828996 0

original_glcm_MaximumProbability 0.089125606 0.089125606 0

original_glcm_SumAverage 40.09024968 40.09024968 0

original_glcm_SumEntropy 4.635501946 4.635501946 0

original_glcm_SumSquares 19.865254 19.865254 0

original_gldm_DependenceEntropy 6.550399892 6.550399892 0

original_gldm_DependenceNonUniformity 120.9761051 120.9761051 0

original_gldm_DependenceNonUniformityNormalized 0.144535371 0.144535371 0

original_gldm_DependenceVariance 18.00291477 18.00291477 0

original_gldm_GrayLevelNonUniformity 62.29510155 62.29510155 0

original_gldm_GrayLevelVariance 24.73367791 24.73367791 0

original_gldm_HighGrayLevelEmphasis 383.9199522 383.9199522 0

original_gldm_LargeDependenceEmphasis 37.44922342 37.44922342 0

original_gldm_LargeDependenceHighGrayLevelEmphasis 20425.03584 20425.03584 0

original_gldm_LargeDependenceLowGrayLevelEmphasis 0.075562484 0.075562484 0

original_gldm_LowGrayLevelEmphasis 0.005605862 0.005605862 0

original_gldm_SmallDependenceEmphasis 0.318186003 0.318186003 0

original_gldm_SmallDependenceHighGrayLevelEmphasis 84.05116859 84.05116859 0

original_gldm_SmallDependenceLowGrayLevelEmphasis 0.003630175 0.003630175 0

original_glrlm_GrayLevelNonUniformity 48.265238 48.265238 0

original_glrlm_GrayLevelNonUniformityNormalized 0.066018368 0.066018368 0

original_glrlm_GrayLevelVariance 24.66124095 24.66124095 0

original_glrlm_HighGrayLevelRunEmphasis 362.3993952 362.3993952 0

original_glrlm_LongRunEmphasis 1.756790194 1.756790194 0

original_glrlm_LongRunHighGrayLevelEmphasis 758.781125 758.781125 0

original_glrlm_LongRunLowGrayLevelEmphasis 0.007773425 0.007773425 0

original_glrlm_LowGrayLevelRunEmphasis 0.006164929 0.006164929 0

original_glrlm_RunEntropy 4.555631762 4.555631762 0

original_glrlm_RunLengthNonUniformity 602.3643647 602.3643647 0

original_glrlm_RunLengthNonUniformityNormalized 0.819694212 0.819694212 0

original_glrlm_RunPercentage 0.868853966 0.868853966 0

original_glrlm_RunVariance 0.382061813 0.382061813 0

original_glrlm_ShortRunEmphasis 0.920028572 0.920028572 0

original_glrlm_ShortRunHighGrayLevelEmphasis 322.2128305 322.2128305 0

original_glrlm_ShortRunLowGrayLevelEmphasis 0.005979334 0.005979334 0

original_glszm_GrayLevelNonUniformity 18.29530201 18.29530201 0

original_glszm_GrayLevelNonUniformityNormalized 0.061393631 0.061393631 0

original_glszm_GrayLevelVariance 21.55781271 21.55781271 0

original_glszm_HighGrayLevelZoneEmphasis 262.7449664 262.7449664 0

original_glszm_LargeAreaEmphasis 93.21812081 93.21812081 0

original_glszm_LargeAreaHighGrayLevelEmphasis 51136.9698 51136.9698 0

original_glszm_LargeAreaLowGrayLevelEmphasis 0.184875307 0.184875307 0

original_glszm_LowGrayLevelZoneEmphasis 0.010725736 0.010725736 0

original_glszm_SizeZoneNonUniformity 138.7248322 138.7248322 0

original_glszm_SizeZoneNonUniformityNormalized 0.465519571 0.465519571 0

original_glszm_SmallAreaEmphasis 0.709478222 0.709478222 0

original_glszm_SmallAreaHighGrayLevelEmphasis 170.3079618 170.3079618 0

original_glszm_SmallAreaLowGrayLevelEmphasis 0.009509532 0.009509532 0

original_glszm_ZoneEntropy 5.514483642 5.514483642 0

original_glszm_ZonePercentage 0.356033453 0.356033453 0

original_glszm_ZoneVariance 85.32918562 85.32918562 0

original_ngtdm_Busyness 0.19930534 0.19930534 0

original_ngtdm_Coarseness 0.008985148 0.008985148 0

original_ngtdm_Complexity 617.8974921 617.8974921 0

original_ngtdm_Contrast 0.183769662 0.183769662 0

original_ngtdm_Strength 2.786428904 2.786428904 0

Reviewer #4: In the manuscript

„Adding the temporal domain to PET radiomic features”,

We would like to thank the reviewer for this positive feedback.

- How were the blood time-activity concentration curves estimated, by blood-sampling or image-based; if so, was the left ventricle used?

We would like to thank the reviewer for this suggestion, we indeed forgot to mention this in the manuscript. The image-derived input function was based on a 10 mL VOI of the descending aorta on which endothelial wall and calcifications were excluded to identify only blood, drawn on the images obtained during the first 60 seconds. We added this to S1, which contains more information on image acquisition and reconstruction.

- The resolution of figure 4 is way to low, it is hardly possible to see anything. Pleas also include the risk tables below the Kaplan-Meier curves.

We thank the reviewer for pointing this out, we have resolved this issue by using a larger font size and larger images. Also, the survival curves have been moved to an additional supporting information file (S3), combined with the association of radiomic features with other clinical parameters.

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

Reviewer #4: No

Attachment

Submitted filename: 20200720 point by point response temporal PET radiomics.docx

Click here for additional data file.^{(69.9KB, docx)}

PLoS One. doi: 10.1371/journal.pone.0239438.r003

Decision Letter 1

Jason Chia-Hsun Hsieh

7 Sep 2020

Adding the temporal domain to PET radiomic features

PONE-D-20-05431R1

Dear Dr. Noortman,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Jason Chia-Hsun Hsieh, M.D. Ph.D

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Most of the questions were answered adequately.

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #3: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

Reviewer #3: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #3: N/A

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

Reviewer #3: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

Reviewer #3: Yes

**********

6. Review Comments to the Author

Reviewer #3: The authors have addressed all concerns/comments. I have no further comments.

One minor issue: will we ever have enough dynamic FDG studies to assess the clinical value for dynamic radiomics? Maybe a comment in line with recommendation of Rich Carson to do a dynamic whole body scan for every first patient of the day could be made. The first hour pi (uptake time) is otherwise not used anyway.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #3: No

PLoS One. doi: 10.1371/journal.pone.0239438.r004

Acceptance letter

Jason Chia-Hsun Hsieh

15 Sep 2020

PONE-D-20-05431R1

Adding the temporal domain to PET radiomic features

Dear Dr. Noortman:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Jason Chia-Hsun Hsieh

Academic Editor

PLOS ONE

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

S1 File. Image biomarker standardisation initiative reporting guidelines.

(DOCX)

Click here for additional data file.^{(51.1KB, docx)}

S1 Data. Data file radiomic features and clinical data.

(XLSX)

Click here for additional data file.^{(146KB, xlsx)}

S1 Fig. Estimated Kaplan-Meier overall survival curves for selected features dichotomized at the median, compared using log-rank statistics.

(DOCX)

Click here for additional data file.^{(1.2MB, docx)}

S1 Table. Association between histopathological characteristics and radiomic features calculated using the Mann-Whitney U test or independent samples t-test, after testing for (log-)normality.

(DOCX)

Click here for additional data file.^{(18.8KB, docx)}

Attachment

Submitted filename: 20200720 point by point response temporal PET radiomics.docx

Click here for additional data file.^{(69.9KB, docx)}

Data Availability Statement

All relevant data are within the manuscript and its Supporting Information files.

[pone.0239438.ref001] 1.Parmar C, Grossmann P, Bussink J, Lambin P, Aerts HJWL. Machine Learning methods for Quantitative Radiomic Biomarkers. Sci Rep. 2015;5:13087 10.1038/srep13087 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref002] 2.Limkin EJ, Sun R, Dercle L, Zacharaki EI, Robert C, Reuze S, et al. Promises and challenges for the implementation of computational medical imaging (radiomics) in oncology. Ann Oncol. 2017;28(6):1191–206. 10.1093/annonc/mdx034 [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref003] 3.Hatt M, Tixier F, Pierce L, Kinahan PE, Le Rest CC, Visvikis D. Characterization of PET/CT images using texture analysis: the past, the present… any future? Eur J Nucl Med Mol Imaging. 2017;44(1):151–65. 10.1007/s00259-016-3427-0 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref004] 4.Junttila MR, de Sauvage FJ. Influence of tumour micro-environment heterogeneity on therapeutic response. Nature. 2013;501:346 10.1038/nature12626 [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref005] 5.Kassner A, Thornhill RE. Texture Analysis: A Review of Neurologic MR Imaging Applications. AJNR Am J Neuroradiol. 2010;31(5):809–16. 10.3174/ajnr.A2061 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref006] 6.Tixier F, Vriens D, Cheze-Le Rest C, Hatt M, Disselhorst JA, Oyen WJ, et al. Comparison of Tumor Uptake Heterogeneity Characterization Between Static and Parametric 18F-FDG PET Images in Non-Small Cell Lung Cancer. J Nucl Med. 2016;57(7):1033–9. 10.2967/jnumed.115.166918 [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref007] 7.Woods BJ, Clymer BD, Kurc T, Heverhagen JT, Stevens R, Orsdemir A, et al. Malignant-lesion segmentation using 4D co-occurrence texture analysis applied to dynamic contrast-enhanced magnetic resonance breast image data. J Magn Reson Imaging. 2007;25(3):495–501. 10.1002/jmri.20837 [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref008] 8.Hu Y, Carmona J, Murphy RF, editors. Application of temporal texture features to automated analysis of protein subcellular locations in time series fluorescence microscope images 2006 IEEE International Symposium on Biomedical Imaging; 2006: IEEE. [Google Scholar]

[pone.0239438.ref009] 9.Haralick RM, Shanmugam K, Dinstein I. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics. 1973;SMC-3(6):610–21. [Google Scholar]

[pone.0239438.ref010] 10.Meijer TWH, de Geus-Oei LF, Visser EP, Oyen WJG, Looijen-Salamon MG, Visvikis D, et al. Tumor Delineation and Quantitative Assessment of Glucose Metabolic Rate within Histologic Subtypes of Non-Small Cell Lung Cancer by Using Dynamic (18)F Fluorodeoxyglucose PET. Radiology. 2017;283(2):547–59. 10.1148/radiol.2016160329 [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref011] 11.Hatt M, Majdoub M, Vallieres M, Tixier F, Le Rest CC, Groheux D, et al. 18F-FDG PET uptake characterization through texture analysis: investigating the complementary nature of heterogeneity and functional tumor volume in a multi-cancer site patient cohort. J Nucl Med. 2015;56(1):38–44. 10.2967/jnumed.114.144055 [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref012] 12.Boellaard R, O'Doherty MJ, Weber WA, Mottaghy FM, Lonsdale MN, Stroobants SG, et al. FDG PET and PET/CT: EANM procedure guidelines for tumour PET imaging: version 1.0. Eur J Nucl Med Mol Imaging. 2010;37(1):181–200. 10.1007/s00259-009-1297-4 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref013] 13.Vriens D, Visser EP, de Geus-Oei LF, Oyen WJ. Methodological considerations in quantification of oncological FDG PET studies. Eur J Nucl Med Mol Imaging. 2010;37(7):1408–25. 10.1007/s00259-009-1306-7 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref014] 14.Hatt M, Cheze le Rest C, Turzo A, Roux C, Visvikis D. A fuzzy locally adaptive Bayesian segmentation approach for volume determination in PET. IEEE Trans Med Imaging. 2009;28(6):881–93. 10.1109/TMI.2008.2012036 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref015] 15.Zwanenburg A, Leger S, Vallieres M, Lock S. Image biomarker standardisation initiative—feature definitions v11. CoRR. 2019;1612.07003. [Google Scholar]

[pone.0239438.ref016] 16.Leijenaar RT, Nalbantov G, Carvalho S, van Elmpt WJ, Troost EG, Boellaard R, et al. The effect of SUV discretization in quantitative FDG-PET Radiomics: the need for standardized methodology in tumor texture analysis. Sci Rep. 2015;5:11075 10.1038/srep11075 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref017] 17.Yip SS, Aerts HJ. Applications and limitations of radiomics. Phys Med Biol. 2016;61(13):R150–66. 10.1088/0031-9155/61/13/R150 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref018] 18.van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. 2017;77(21):e104–e7. 10.1158/0008-5472.CAN-17-0339 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref019] 19.Bianco-Martinez E, Baptista MS. Space-time nature of causality. Chaos: An Interdisciplinary Journal of Nonlinear Science. 2018;28(7). [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref020] 20.Goeman JJ, van de Geer SA, de Kort F, van Houwelingen HC. A global test for groups of genes: testing association with a clinical outcome. Bioinformatics. 2004;20(1):93–9. 10.1093/bioinformatics/btg382 [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref021] 21.Collarino A, Garganese G, Fragomeni SM, Pereira Arias-Bouda LM, Ieria FP, Boellaard R, et al. Radiomics in vulvar cancer: first clinical experience using (18)F-FDG PET/CT images. J Nucl Med. 2019;60(2):199–206. [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref022] 22.Liu J, Dong M, Sun X, Li W, Xing L, Yu J. Prognostic Value of 18F-FDG PET/CT in Surgical Non-Small Cell Lung Cancer: A Meta-Analysis. PLoS One. 2016;11(1):e0146195 10.1371/journal.pone.0146195 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0239438.ref023] 23.Ohtaka K, Hida Y, Kaga K, Okamoto S, Shiga T, Tamaki N, et al. Outcome Analysis of 18F-Fluorodeoxyglucose Positron–Emission Tomography in Patients with Lung Cancer After Partial Volume Correction. 2013;33(11):5193–8. [PubMed] [Google Scholar]

[pone.0239438.ref024] 24.Razzak MI, Naz S, Zaib A. Deep Learning for Medical Image Processing: Overview, Challenges and the Future. In: Dey N, Ashour AS, Borra S, editors. Classification in BioApps: Automation of Decision Making. Cham: Springer International Publishing; 2018. p. 323–50. [Google Scholar]

[pone.0239438.ref025] 25.Buvat I ORLHAC F. The dark side of radiomics: on the paramount importance of publishing negative results. J Nucl Med. 2019. [DOI] [PubMed] [Google Scholar]

[pone.0239438.ref026] 26.Cherry SR, Jones T, Karp JS, Qi J, Moses WW, Badawi RD. Total-Body PET: Maximizing Sensitivity to Create New Opportunities for Clinical Research and Patient Care. J Nucl Med. 2018;59(1):3–12. 10.2967/jnumed.116.184028 [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Adding the temporal domain to PET radiomic features

Wyanne A Noortman

Dennis Vriens

Cornelis H Slump

Johan Bussink

Tineke W H Meijer

Lioe-Fee de Geus-Oei

Floris H P van Velden

Roles

Abstract

Background

Methods

Results

Conclusion

Introduction

Materials and methods

Patients and clinical follow-up

Table 1. Clinical characteristics of 35 non-small cell lung carcinoma (NSCLC) lesions in 34 patients.

Patient preparation, data acquisition, image reconstruction and parametrisation

Fig 1. Dynamic [18F]FDG PET acquisition of a patient with non-small cell lung carcinoma.

Image analysis

Volumes of interest

Interpolation and discretisation

Static and parametric features

Dynamic features

Statistical analysis

Results

Bin widths

Differences between scanners

Parametric features

Dynamic features

Fig 2. Spearman correlation matrix of static radiomic features (x-axis) and dynamic texture features (y-axis).

Clinical relevance

Discussion

Conclusion

Supporting information

Acknowledgments

List of abbreviations

Data Availability

Funding Statement

References

Decision Letter 0

Jason Chia-Hsun Hsieh

Roles

Author response to Decision Letter 0

Decision Letter 1

Jason Chia-Hsun Hsieh

Roles

Acceptance letter

Jason Chia-Hsun Hsieh

Roles

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases

Fig 1. Dynamic [¹⁸F]FDG PET acquisition of a patient with non-small cell lung carcinoma.