The choice of statistical methods for comparisons of dosimetric data in radiotherapy
 Abdulhamid Chaikh^{1, 2}Email author,
 JeanYves Giraud^{1, 2},
 Emmanuel Perrin^{3},
 JeanPierre Bresciani^{4, 5} and
 Jacques Balosso^{1, 2}
https://doi.org/10.1186/1748717X9205
© Chaikh et al.; licensee BioMed Central Ltd. 2014
Received: 20 May 2014
Accepted: 2 September 2014
Published: 18 September 2014
Abstract
Purpose
Novel irradiation techniques are continuously introduced in radiotherapy to optimize the accuracy, the security and the clinical outcome of treatments. These changes could raise the question of discontinuity in dosimetric presentation and the subsequent need for practice adjustments in case of significant modifications. This study proposes a comprehensive approach to compare different techniques and tests whether their respective dose calculation algorithms give rise to statistically significant differences in the treatment doses for the patient.
Methods
Statistical investigation principles are presented in the framework of a clinical example based on 62 fields of radiotherapy for lung cancer. The delivered doses in monitor units were calculated using three different dose calculation methods: the reference method accounts the dose without tissues density corrections using Pencil Beam Convolution (PBC) algorithm, whereas new methods calculate the dose with tissues density correction for 1D and 3D using Modified Batho (MB) method and Equivalent Tissue air ratio (ETAR) method, respectively. The normality of the data and the homogeneity of variance between groups were tested using ShapiroWilks and Levene test, respectively, then nonparametric statistical tests were performed. Specifically, the dose means estimated by the different calculation methods were compared using Friedman’s test and Wilcoxon signedrank test. In addition, the correlation between the doses calculated by the three methods was assessed using Spearman’s rank and Kendall’s rank tests.
Results
The Friedman’s test showed a significant effect on the calculation method for the delivered dose of lung cancer patients (p <0.001). The density correction methods yielded to lower doses as compared to PBC by on average (−5 ± 4.4 SD) for MB and (−4.7 ± 5 SD) for ETAR. Posthoc Wilcoxon signedrank test of paired comparisons indicated that the delivered dose was significantly reduced using densitycorrected methods as compared to the reference method. Spearman’s and Kendall’s rank tests indicated a positive correlation between the doses calculated with the different methods.
Conclusion
This paper illustrates and justifies the use of statistical tests and graphical representations for dosimetric comparisons in radiotherapy. The statistical analysis shows the significance of dose differences resulting from two or more techniques in radiotherapy.
Keywords
Background
Radiotherapy techniques
The main challenge in radiation therapy for cancer treatment is to obtain the highest probability of tumor control or cure with the least amount of morbidity and toxicity to normal surrounding tissues (organs at risk). Currently, numerous different machines and several techniques are used to irradiate the tumors, as three–dimensional radiation therapy (3DRT), intensitymodulated radiation therapy (IMRT), tomotherapy, particle therapy and volumetricmodulated arc therapy (VMAT). The expected clinical results of radiotherapy are related to the calculated dose. The advance in technology provides successive generations of Treatment Planning Systems (TPS) for radiotherapy which include new dose calculation algorithms and allow new irradiation techniques. These algorithms compute the dose for a given technique, subsequently showing the results as dosimetric parameters and displaying dose volume histograms or spatial isodoses. The validation of the treatment plan in radiotherapy is based on the assessment of these TPS output. When a new technique or a new algorithm is implemented, the calculated and distributed doses can differ from those computed by algorithms that constitute the current standard and reference. It is very easy to compare the spatial dose distribution of two DICOM RT files (the file produced by the reference method and the file produced by the new method) by using γ index or χ index [1, 2]. Any difference between the computed dose will be visualized using γ or χ maps [3]. If ignored, this alteration could endanger the clinical outcome of the treatment. In particular, the heterogeneity correction introduces an under dosage to the target when the latter is imbedded in low density tissues, as in thoracic situation. Therefore, prescription habits should be adapted to the new calculation methods and strong connection between dosimetric methods and prescription understanding should be established to avoid dosimetric improvement to turn into clinical regression [4]. Currently, algorithms and irradiation techniques are both able to modify independently the dosimetric outcome. Although the medical physicists validate any treatment plan according to the international recommendations for radiotherapy, these alterations are supposed to have consequences on the clinical outcome [5, 6]. One should provide the radiation oncologist with a tool allowing him to assess the reality and the extent of these modifications, to help him integrate them into his everyday practice. Ideally, the assessment should be possible using small data sets of patients. That way, each department could perform the assessment tests without investing too much time and resources. At this extend, radiotherapy offers the valuable situation in which each patient case could be recalculated in many different ways easily providing paired series of data. This paper presents a series of statistical tests implemented in a step by step procedure that should help the radiation oncologist comparing the dosimetric outcome of different dose calculation algorithms. The procedure is presented using a concrete example and putting the emphasis on the application in radiotherapy rather than the underlying mathematical principles, which are not detailed. Each section gives a brief description of the aim and the condition of use of the statistical tests. The goal is to provide the radiation oncologist and the physicists interpretable results to help them to validate new dose calculation methods.
The need of statistical analysis in radiotherapy
When comparing algorithms or techniques, the statistical analysis has to answer the key question: are there real differences between two or more calculated dose distributions? By principle, the main objective of any progress in radiotherapy is to improve the clinical outcome by increasing the dose to the tumor and reducing the dose to the organs at risk. The dosimetric progress, by the improvement of the accuracy of the calculation, is supposed to contribute to this goal. Statistical tests are used to make probabilitybased decisions regarding differences measured between different conditions, e.g., different radiation techniques. For example, are there real dosimetric differences, assessed by the physicist, which could be anticipated by the radiation oncologist as a real benefit or a possible risk for the patient outcome using RT3D or IMRT techniques.
To compare different techniques or dose calculation algorithms, all dosimetric data are calculated with a unique set of images for a given patient, whatever the number of different algorithms to compare. This provides the radiotherapist with paired data that can be analyzed using statistical tests for repeated measures, e.g., paired ttest when only comparing two techniques or ANOVA when comparing more than two techniques. All statistical tests used in this study were performed with the R programming language [7, 8].
Methods
Data collection
The first step consists in preparing and arranging data for further analysis. In order to evaluate the treatment plan for radiotherapy, several dosimetric parameters can be considered:

Delivered dose in monitor unit.

Spatial distribution of dose: for example the isodose curves 100% and 95% inside the PTVs can be compared.

Volumetric distribution: the use of dose volume histogram allows comparing the maximum dose, minimum dose, mean dose, etc.

Quality index: to compare the conformity plan, the dose homogeneity in PTV and the protection of organs at risk.
Data collection table for statistical analysis
Patients  Reference  Tested method 1  Tested method 2 

1  D_{r,1}  D_{1,1}  D_{2,1} 
2  D_{r,2}  D_{1,2}  D_{2,2} 
n  D_{r,n}  D_{1,n}  D_{2,n} 
μ_{r} ± SDr  μ_{1} ± SD_{1}  μ_{2} ± SD_{2} 
A positive value means that the dosescalculated by the tested method are higher than the doses calculated by the reference method (D_{tested} > D_{reference}). A negative value would mean the opposite (D_{tested} < D_{reference}).
Assuming that α = 5% the Zα = 1.96 shows the critical value for 95% confidence level for normal distribution.
To determine the sample size, one should beforehand, set the significance level (α) and the statistical power of the test to come. α defines the probability of “erroneously” concluding that the observed difference between means reflects a real difference between methods when this difference between means was actually observed by chance. α is typically set to 0.05 (pvalue) which corresponds to a 5% chance to conclude to a significant difference when there is no actual difference. Statistical power corresponds to the probability of detecting an actual difference. Applied to the comparison of radiotherapy methods, statistical power therefore corresponds to the probability of “correctly” concluding that the observed difference between means reflects a real difference between methods. A conventional choice of power is either 80% or 90% [9, 10]. Power is usually calculated before data collection in order to estimate the required sample size. In some case, statistical power is calculated after data collection to verify whether the nonsignificance of the result might be due to a lack of power, but this practice is usually discouraged [11]. Here we calculated statistical power after data collection for illustrative purposes. For that, we used PS software, which provides estimates of power and sample size, and allowed us to plot sample size versus power for a specified alternative hypothesis [12].
In addition to this power calculation, we used bootstrapping to estimate the minimum number of patients or cases that we would have needed to observe a significant difference between methods or groups with our data. We chose this a posteriori approach because we already had data collected for 62 beams. It also allowed us to compare the results with the estimation provided by the power analysis.
As an additional note, we would like to mention that Pvalues do not really measure the importance of the effect. Therefore, statistical tests can be complemented by an assessment of the effect size (effsize). We choose to use “Pearson’s r”, to measure and estimate the amount of total variance in the data resulting from the difference between the compared groups.
Comparing methods
Several statistical tests can be used to compare means. These tests can be classified in two categories: parametric and nonparametric tests. Parametric tests make assumptions about the type and distribution of the data. If the data fulfill the assumptions of parametric data, parametric tests can be used. If the data do not fulfill these assumptions, nonparametric tests must be used. Nonparametric tests make fewer assumptions about the type and distribution of the data.
Checking assumptions of parametric data
Comparison of techniques
To compare two techniques (i.e., two means) when measures are repeated, paired Student parametric test or Wilcoxon signedrank nonparametric test can be used, as shown in Figures 2 and 3. With independent measures (i.e., independent samples), independent Student ttest and Mann Whitney test for nonparametric data are usually used.
To compare more than two techniques, a twostep analysis should be conducted. The first step consists in conducting an omnibus test to assess whether there is an overall effect of technique, i.e., whether all techniques are the same or not. If the data fill the assumptions of parametric data (normally distributed data and homogeneity of variance between groups), this omnibus test can be performed using a oneway Analysis of Variance (ANOVA) for repeated measures, with correction for violation of sphericity if needed [15]. To compare the outcome of different techniques for more than one dosimetric parameter, one can use a multivariate analysis of variance (MANOVA), but this point will not be discussed here. If the data do not fill the assumptions of parametric data, the omnibus test must be performed using a nonparametric equivalent of the repeated measures ANOVA, namely Friedman’s repeated measures test (the KruskalWallis test can be used for independent measures). Once this omnibus test has been performed, and if a main effect has been detected, the second step consists in conducting multiple pairedttests (multiple comparisons) to compare techniques in a two by two fashion. With parametric data, these comparisons can be performed using multiple paired ttests (see above). With nonparametric data, multiple Wilcoxon signedrank tests can be used. For both parametric and nonparametric multiple comparisons, the significance level must be corrected in order to avoid inflating type1 error probability, i.e., the probability of detecting a significant difference when actually there is not any (i.e., false positive). One way of doing that is to divide the significance threshold by the total number of paired comparisons performed (e.g., with three comparisons, alpha = 0.05/3 = 0.0166). This correction method is known as Bonferroni correction. Bonferroni correction is the most conservative correction method. Specifically, it controls strictly the probability of false positive, but this is at the cost of reducing the statistical power of the test, i.e., the ability to detect an actual difference. When power is an issue, less conservative correction methods can be used, as for instance the Holm method, in which the correction factor applied depends not only on the number of comparisons performed but also on the rank of the p values. Finally, several other posthoc procedures are available to compare several means, as Tukey’s honest significant difference test (HSD), NewmanKeuls test, or Fisher’s least significant difference (LSD) test, to cite a few. Tukey’s HSD is one of the most widely used, especially when the number of means to compare is relatively large, as it offers good control over false positives while keeping decent power (see [16] for a specific comparison between posthoc procedures).
Correlation between techniques
To measure the strength of the relationship between two techniques (i.e., estimate how much two techniques are related), for parametric data, Pearson correlation coefficient can be used. For nonparametric data, one should rather use Spearman’s ‘rho’ correlation coefficient or Kendall’s ‘tau’. In this study, the data did not fill the assumptions of parametric data. Therefore, Spearman’s rank correlation and Kendall’s rank were used.
Medical decision
A pvalue (p ≥ 0.05) implies that the differences observed between means probably occurred by chance (rather than reflecting a real difference between the methods). In this case, we can assume that there is not enough evidence to conclude that the new technique gives rise to significant dosimetric difference with respect to the reference technique. Conversely, a p < 0.05 implies that the differences observed probably reflect existing differences between the methods. In this case, we could expect a significant dosimetric difference and therefore some medical impact.
Clinical tests for radiotherapy
Characteristics of the patients treated with 3D radiotherapy using 18MV, location tumour, dose prescription and treatment fields
Patients  Locations  Dose (Gy)  Fields 

1  Lung parenchyma  66  6 
2  Left retro cardiac  66  17 
3  Top left lung  70  10 
4  Mediastina  60  9 
5  Mediastina  60  12 
6  Oesophagus  55  8 
The aim was to compare the delivered dose in monitor units resulting from three dose calculation methods keeping exactly the same beam setting. Calculations dose were performed, for the present demonstration, using the Pencil Beam Convolution (PBC) algorithm integrated in TPS Eclipse® (Varian, version 8.1). PBC algorithm includes two calculation modes: without heterogeneity correction using (PBC) and with heterogeneity correction using the Modified Batho (MB) method and Equivalent Tissue air ratio (ETAR) method. Heterogeneity corrections are always based on relative electron densities obtained from a CTscan [17, 18].
Reference method: calculates the dose without taking into account the tissues densities using PBC.
Where μ_{m} and μ_{w} are the linear attenuation coefficients of the material in layer (m) and water (w) respectively; Z_{bu} is the buildup depth and Z_{m} is the distance along the beam from the surface to the layer (m) in the phantom. μ_{en}/ρ is the mass energy absorption coefficient of the material in layer (N).
Where ${d}^{\prime},\tilde{r}$ are the effective values of depth (d) and beam radius (r) respectively.
The statistical analyses in this study were performed to assess whether using new method MB or ETAR rather than the reference method (PBC) would give rise to a significant reduction or increase in the delivered dose. The analyses were carried out using the R environment for statistical computing and visualization. All data were imported in R from Excel®.
Results
Assessment of the assumptions of parametric data
Normality
Observed results from ShapiroWilks test, skewness and excess kurtosis for the reference and two new methods
Tests  Results  Reference  Method 1  Method 2 

ShapiroWilks  W  0.80  0.81  0.81 
pvalue  <0.001  < 0.001  < 0.001  
Skewness  Skewness/standard error  6.05  5.64  5.68 
pvalue  < 0.001  < 0.001  < 0.001  
Excess Kurtosis  Kurtosis/standard error  6.15  5.04  5.14 
p value  < 0.001  < 0.001  < 0.001 
Homogeneity of variance
Comparisons of dose calculation methods
Therefore, one should assess whether this difference is significant or rather occurred by chance. Friedman rank sum test showed a significant overall effect of technique,χ^{2}(2) = 53.45, p < 0.001.
Relationship between dose calculation methods using correlation and regression
The 95% BCa for the slope was [0.90 ; 0.96] and [0.92; 0.97] for the comparison between PBC with MB and ETAR respectively. The 95% BCa for the intercept was [−0.83 ; 3.67] and [−1.14 ; 2.95] for the comparison between PBC with MB and ETAR respectively. In both cases, the coefficient of determination was very high (R^{2} = 0.996) with p < 0.001. The 95% confidence bootstrap percentile (95% BCa) interval for correlation and regression has been computed using nonparametric bootstrapping with 2000 replicates.
Medical decision
The mean comparison tests between the PBC method with MB or ETAR indicated significant differences in dose calculation (p < 0.001). In other words, the observed differences probably reflect existing differences between the methods. In addition, the bootstrapping procedure indicated that significant differences between the reference and the new methods could be observed with as little as 8 and 10 beams, respectively. Therefore, the observed dosimetric differences between methods very likely bear a clinical impact.The regression analyses show slopes of 0.94 and 0.95 for MB and ETAR, respectively (see Figure 9). The slope represents the change associated with the MB and ETAR compared to PBC. Should these methods be equivalent, these values for slope would have been equal to one. In this study, the slope values are below 1, with 95% confidence intervals of [0.90; 0.96] and [0.92; 0.97], respectively. These slope values confirm the average difference of 5% and 4.7% computed with equation 1 for MB and ETAR. This overall difference is the main result of this study, and confirms that the prescribed dose should be adjusted by on average +5% and +4.7% using MB and ETAR, respectively.
However, different cancer sites should be considered individually before making any general rule of dose modification. We recommend presenting the complete set of statistical information including mean, SD, confidence intervals, pvalue, sample size and graphical data analysis to the medical staff. These information will help the radiationoncologists to take a decision about the modification of the irradiation technique and dose prescription. The confidence interval, which is routinely computed by all statistical packages, shows the size of difference which could be observed. The width of a confidence interval depends on the mean and standard deviation of the results.
Discussion
Among the numerous statistical methods available, the medical physicist has to make a choice well adapted to the particularities of radiation therapy. At first, the particular nature of the data and the way they are produced need a deep analysis of their quality. In particular, one should assess whether the data at hand fulfill the assumptions of parametric data, i.e., are distributed normally and have similar variance between groups. When the data fulfill these assumptions, Student ttest or one way ANOVA can be used to compare means. But if data do not fulfill these assumptions, alternative nonparametric tests should be used, as for instance the Wilcoxon signedrank test or Friedman ANOVA (when dealing with repeated measures). The nonparametric Wilcoxon rank test takes into account the signedrank of the difference between each pair of measures instead of using all the absolute data. It does not require a normal distribution and does not considers the size of the difference. These features make the Wilcoxon signedrank test particularly fit for radiotherapy data analysis, since these data are “naturally” paired according to the possibility to generate multiple different results with each medical case. In case of simultaneous comparison of multiple parameters (i.e., not only the delivered doses per monitor unit but also volumetric distribution, etc.), multivariate ANOVA can be performed (not used in this paper). However, an indepth discussion between medical physicists and statisticians is strongly encouraged in order to determine which tests and analyses are the most appropriate for the data at hand (e.g., use of parametric vs nonparametric tests).
We compared one parameter using three methods for calculation dose. The example given in this study concerning the delivered dose using 1D or 3D dose calculation is an exemplary problem in radiotherapy. The clinical effect in radiotherapy depends on the delivered dose, a small difference in delivered dose being worth to be considered since it could affect the clinical results. Concerning the delivered dose, we observed a significant difference between reference method and both tested methods (p < 0.05). However, pvalue gives little information about the results, for example, using 5% level for alpha error, the pvalue shows only if the difference between the two techniques has a statistical significance or not. It does not give the size of the difference between two techniques. This information can be gathered computing the percentage of difference between two techniques (as we did here) but also computing the effect size (here the Pearson’s r) which is a standardized measure (always between −1 and 1) of the importance of the observed effect. In our study, the measured effect sizes were quite large (above 0.5 in absolute value) for both comparisons of the new methods with the reference method.
The calculation of sample size is crucial in any clinical study. It depends on the significance level, the wished power, as well as the expected effect size and SD. The sample size for clinical and statistical studies is the main difficulty for radiotherapy. For practical reasons, it would be welcome to use only few patients for realizing the statistical analysis, and then to generalize the results to a large population. In radiotherapy it is rare to use a large number of patients in order to validate the novel irradiation technique at the level of a common department. Most of the studies in radiotherapy are including between 10 and a few hundred patients. Most of the time, a novel technique is quickly integrated in a radiotherapy department and the physicists has not enough time to test a large sample of patient. In this case, the radiationoncologist and the physicists will have to realize the statistical analysis using a small sample size, which is actually possible providing one uses cautiously the proper methods. As mentioned above, one possibility to estimate the required sample size for an expected effect size is to use a power analysis. Because we already had data collected (62 beams), we chose to rather use a bootstrap approach to estimate the minimum number of patients or cases that we would have needed to observe a significance difference between methods or groups. We have seen that a significant difference between our methods could be observed with relatively small sample, namely eight and ten beams.
Interestingly enough, this approach is basically funded on the analysis of differences and in particular dose distribution differences. Therefore it can be adapted to compare any type of situation resulting at last in dosimetric differences as different irradiation methods, different machines, even assessing the identity of mirror machines, and so on.
Conclusion
We illustrate in this study the use of statistical tests for paired observations well adapted to the specificity of radiotherapy where different data sets can be obtained of the same patient. To compare two radiotherapy techniques, the standard ttest can be used with parametric data, whereas nonparametric Wilcoxon signedrank test should be used if the dose differences do not fulfill the assumptions of parametric data. To compare several techniques, the standard analysis of variance (ANOVA) can be used with parametric data, whereas nonparametric tests as the Friedman ANOVA should be used with nonparametric data. Here we suggested statistical methods to compare calculated doses in radiotherapy. These methods can also be used to compare measured doses, e.g., using two or more dosimeters, as well as to compare calculated and measured doses.
Declarations
Acknowledgement
The authors acknowledge the financial support of French National Research Agency (ANR11TECS018) for research funding.
Authors’ Affiliations
References
 Low DA, Harms WB, Mutic S, Purdy JA: A technique for the quantitative evaluation of dose distributions. Med Phys 1998, 25: 656.View ArticlePubMedGoogle Scholar
 Bakai A, Alber M, Nusslin F: A revision of the gammaevaluation concept for the comparison of dose distributions. Phys Med Biol 2003, 48: 35433553.View ArticlePubMedGoogle Scholar
 Chaikh A, Giraud JY, Balosso J: A 3D quantitative evaluation for assessing the changes of treatment planning system and irradiation techniques in radiotherapy. Int J Cancer Ther Oncol 2014,2(3):02033.View ArticleGoogle Scholar
 Chaikh A, Giraud JY, Balosso J: A method to quantify and assess the dosimetric and clinical impact resulting from the heterogeneity correction in radiotherapy for lung cancer. Int J Cancer Ther Oncol 2014,2(1):020110.View ArticleGoogle Scholar
 ICRU Report No.50: Prescribing, Recording and Reporting Photon Beam Therapy. Bethesda, Maryland: International Commission on Radiation Units and Measurements; 1993.Google Scholar
 ICRU Report No.62: Prescribing, Recording and Reporting Photon Beam Therapy Supplement to ICRU Report 50. Bethesda, Maryland: International Commission on Radiation Units and Measurements; 1999.Google Scholar
 de Marques SJP: Applied Statistics Using SPSS, STATISTICA, MATLAB and R. 2nd edition. Berlin: Springer; 2007.View ArticleGoogle Scholar
 The R project for statistical computing http://www.rproject.org
 Bolton S, Bon C: Practical and Clinical Applications. Volume 135. 4th edition. Newyork Dekker: Marcel, Inc; 2004.Google Scholar
 Suresh KP, Chandrashekara S: Sample size estimation and power analysis for clinical research studies. J Hum Reprod Sci 2012,5(1):713.PubMed CentralView ArticlePubMedGoogle Scholar
 Hoenig JM, Heisey DM: The abuse of power: the pervasive fallacy of power calculations for data analysis. Am Stat 2001,55(1):16.View ArticleGoogle Scholar
 Power and Sample Size Calculation http://biostat.mc.vanderbilt.edu
 Dufour JM, Farhat A, Gardiol L, Khalaf L: Simulationbased finite sample normality tests in linear regressions. Economet J 1998, 1: 154173.View ArticleGoogle Scholar
 Razali NM, Wah YB: Power comparisons of ShapiroWilk, KolmogorovSmirnov, Lilliefors and AndersonDarling tests. J Stat Mod Analyt 2011,2(1):2133.Google Scholar
 Algina J, Olejnik S: Conducting power analyses for ANOVA and ANCOVA in betweensubjects designs. Eval Health Prof 2003,26(3):288314.View ArticlePubMedGoogle Scholar
 Rodger RS, Roberts M: Comparison of Power for Multiple Comparison Procedures. J Methods Meas Soc Sci 2013,4(1):2047.Google Scholar
 Task Group No. 65, the Radiation Therapy Committee of the American Association of Physicists in Medicine: Tissue inhomogeneity corrections for MV photon beams. Madison, WI: Medical Physics Publishing 2004. https://www.aapm.org/pubs/reports/rpt_85.pdf
 Ahnesjö A, Aspradakis MM: Dose calculations for external photon beams in radiotherapy. Phys Med Biol 1999, 44: 99155.View ArticleGoogle Scholar
 Loftus GR, Masson MEJ: Using confidence interval in withinsubject designs. Psychon Bull Rev 1994,1(4):476490.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate. Please note that comments may be removed without notice if they are flagged by another user or do not comply with our community guidelines.