- Email: [email protected]
Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients
Radiotherapy and Oncology xxx (2015) xxx–xxx
Contents lists available at ScienceDirect
Radiotherapy and Oncology journal homepage: www.thegreenjournal.com
Original article
Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients Uffe Bernchou a,b,⇑, Olfred Hansen a,c, Tine Schytte c, Anders Bertelsen b, Andrew Hope d, Douglas Moseley d, Carsten Brink a,b a d
Institute of Clinical Research, University of Southern Denmark, Odense; b Laboratory of Radiation Physics; Department of Radiation Oncology, Princess Margaret Hospital, Toronto, Canada
a r t i c l e
i n f o
Article history: Received 11 May 2015 Received in revised form 30 June 2015 Accepted 16 July 2015 Available online xxxx Keywords: Lung cancer Density change Computed tomography Cone beam Response modelling
c
Department of Oncology, Odense University Hospital, Denmark ; and
a b s t r a c t Background and purpose: This study investigates the ability of pre-treatment factors and response markers extracted from standard cone-beam computed tomography (CBCT) images to predict the lung density changes induced by radiotherapy for non-small cell lung cancer (NSCLC) patients. Methods and materials: Density changes in follow-up computed tomography scans were evaluated for 135 NSCLC patients treated with radiotherapy. Early response markers were obtained by analysing changes in lung density in CBCT images acquired during the treatment course. The ability of pre-treatment factors and CBCT markers to predict lung density changes induced by radiotherapy was investigated. Results: Age and CBCT markers extracted at 10th, 20th, and 30th treatment fraction significantly predicted lung density changes in a multivariable analysis, and a set of response models based on these parameters were established. The correlation coefficient for the models was 0.35, 0.35, and 0.39, when based on the markers obtained at the 10th, 20th, and 30th fraction, respectively. Conclusions: The study indicates that younger patients without lung tissue reactions early into their treatment course may have minimal radiation induced lung density increase at follow-up. Further investigations are needed to examine the ability of the models to identify patients with low risk of symptomatic toxicity. Ó 2015 Elsevier Ireland Ltd. All rights reserved. Radiotherapy and Oncology xxx (2015) xxx–xxx
Dose escalation for non-small cell lung cancer (NSCLC) is typically limited by increasing risk of severe adverse events such as radiation pneumonitis (RP). Therefore, identifying patients with low or high risk of developing normal tissue toxicity is of importance if the dose is to be escalated safely. Response predictions are typically based on models including dosimetric parameters [1]. However, the predictive power of the models could potentially be increased by adding other baseline clinical factors and predictive markers obtained prior to treatment or response markers acquired early into the treatment course [2]. A recent study has detected normal tissue reactions in terms of density change in healthy lung tissue in cone-beam computed tomography (CBCT) images obtained relatively early during a fractionated radiotherapy (RT) treatment course [3]. Although lung density change not necessarily implies symptomatic toxicity, RP ⇑ Corresponding author at: Laboratory of Radiation Physics, Odense University Hospital, Sdr. Boulevard 29, DK-5000 Odense C, Denmark. E-mail address: [email protected] (U. Bernchou).
typically manifests as increases in density within the healthy lung tissue as observed in follow-up computed tomography (CT) images [4]. In the current study, density changes in follow-up CT were evaluated in a cohort of NSCLC patients treated with CBCT image guided RT. Lung Radioresponsiveness (LuRa) was defined as the slope of a linear fit to dose–effect curves produced by relating changes in regional lung density to regional planned dose as proposed by De Ruysscher et al. [5]. LuRa thus provides an objective, continuous measure of the patients’ individual susceptibility to radiation in terms of density change. The current study investigated the ability of pre-treatment factors and CBCT markers (CBCTM) to predict LuRa. Methods and materials The study includes NSCLC patients treated to a prescribed dose of 60 or 66 Gy in 2-Gy fractions in the period from 2007–2013 at Odense University Hospital. A total of 135 patients (Table 1) met the following inclusion criteria: (1) CBCTs were acquired
http://dx.doi.org/10.1016/j.radonc.2015.07.021 0167-8140/Ó 2015 Elsevier Ireland Ltd. All rights reserved.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
2
Predicting lung density changes after RT
Table 1 Characteristics of the 135 patients. Characteristic
na
%
Gender Male Female
71 64
53 47
Age (years) Median Range
68 33–85
Performance (WHO) 0 1 2 3
40 70 24 1
FEV1 (l/min) Median Range
1.8 0.6–4.5
Stage 1–2 3 4 Recurrence
27 92 5 11
20 68 4 8
Histology Squamous cell carcinoma Adenocarcinoma Otherb
63 54 18
47 40 13
Current smoking status Smoking Not smoking
71 64
53 47
GTV volume (cm3) Median Range
31 0–400
PTV volume (cm3) Median Range Total radiation dose (Gy) 60 66
29 52 18 1
Image registration
385 62–1648 36 99
Mean lung dose (Gy) Median Range
17 2–24
Chemotherapy No Only neoadjuvant Only concomitant Both neoadjuvant and concomitant
30 18 1 86
groups of treatment fractions were analysed: (I) 1, (II) 2–3, (III) 10–13, (IV) 20–23, and (V) 30–33. To be included in the study a patient must have had at least one CBCT acquired in groups (I) and (II) and at least one CBCT in either of the groups (III)–(V), and all CBCTs must be acquired on the same accelerator and using the same imaging technique and filter settings. If several scans were acquired in each group, the earliest scan was used. As described below, density changes in CBCT images were monitored relative to the CBCT density at the 1st fraction, while the scans in group (II) were included to measure the patient specific noise level, and response markers were extracted from the scans in group (III)– (V). The fraction numbers within each group were chosen to match the action limit protocol with a few extra scans included to increase the number of valid scans in the cases where the imaging protocol or the accelerator was changed. Scans belonging to group (II)–(V) will in the following be termed 2nd, 10th, 20th, and 30th fraction, respectively. The number of CBCT scans available was 133, 129, and 113 at the 10th, 20th, and 30th fraction, respectively, and in total 645 CBCT scans were analysed. The clinical scans were either 3D or 4D CBCT. However, for each scan used in the study, the raw projection images were reconstructed to 3D with a slice thickness of 2 mm and pixel size of 1 mm by the Feldkamp–David–Kress [7] algorithm implemented in the open-source cone beam Reconstruction Toolkit [8]. A total of 222 follow-up CT scans acquired within 6 months from commencement of RT were evaluable for the LuRa determination (see details below). The follow-up CT scans were typically acquired in the inspirational breath-hold at a diagnostic CT scanner using a median slice thickness of 5 mm (range 2–7.5 mm) and a pixel size of 0.7 mm (range 0.6–1.0 mm).
27 73
22 13 1 64
Abbreviations: GTV = gross tumour volume; PTV = planned target volume. a n denoted number of patients unless otherwise indicated. b Large cell carcinoma, undifferentiated, atypical carcinoid, unspecified.
throughout the treatment as specified below. (2) Eligible follow-up CT scans were acquired within 6 months from commencement of RT without clinical or radiological evidence of local recurrence or thoracic metastases. (3) Acceptable image registration, as described below, was evaluated on visual validation. (4) The volume of investigated healthy lung tissue was larger than 250 cm3, to ensure sufficient data for the individual patient. (5) No sign of atelectasis was seen in the investigated lung tissue volume during treatment or at follow-up. Details on treatment planning, chemotherapy and follow-up are described elsewhere [6]. Imaging data Treatment planning was based on the mid-ventilation phase of a 4D CT scan with slice thickness of 2.5 or 3 mm and pixel size of 1 mm. RT was delivered on Elekta Synergy accelerators equipped with an X-ray Volume Imaging (XVI) system. CBCT image guided RT (IGRT) was performed either daily or at the first three fractions and at least at the 10th, 20th, and 30th fraction, according to an action limit protocol. Imaging data acquired at the following five
The main steps in assessing density changes in healthy lung tissue from CBCT and follow-up CT image data and relating this to the locally delivered dose have been described in detail elsewhere [3,6,9]. Briefly, deformable image registrations were performed using the elastix freeware toolbox [10]. CBCT-CBCT registrations allow the creation of 3D maps of CBCT density changes throughout the treatment by subtraction of the images acquired at the 1st fraction from the images acquired at subsequent fractions. By registration of the CBCTs to the planning CT, the density changes were related to the locally planned dose. Similarly, the follow-up CT scans were co-registered to the planning CT for the investigation of lung density changes post treatment. For both follow-up CT and CBCT, a 5 5 5 mm wide median filter was applied on the 3D maps of density change in order to suppress noise. Data analysis Density changes were evaluated in the regions of healthy lung tissue within the CBCT field of view. Healthy lung tissue was defined as the total lung delineated during treatment planning, eroded by 10 mm and with the PTV subtracted to avoid boundary effects. Patient individual dose–effect curves at follow-up were created by averaging the density changes post treatment for regions receiving the same planned dose in 5-Gy bins. The LuRa for a patient was defined as the slope of a linear regression to the curve as suggested by De Ruysscher et al. [5]. The high dose regions (>45 Gy) of the curve were excluded from the fit as previous investigations have shown that dose response is non-linear in this regime, at least in the early regime (<3 months) [6]. For patients with more than one evaluable follow-up scan, the highest LuRa value was chosen. Two types of response markers were extracted from the cone beam images: a slope marker (CBCTMslope) and a threshold marker (CBCTMthreshold). CBCTMslope was calculated similar to the LuRa values at follow-up as the slope
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
3
U. Bernchou et al. / Radiotherapy and Oncology xxx (2015) xxx–xxx
Histogram analysis Voxel histograms of the density change relative to the CBCT acquired at the 1st fraction in regions receiving 20–45 Gy planned dose were produced for each available treatment fraction (2nd-30th) for each patient. Although density changes have been observed in the low-dose region in follow-up CT scans [6], the CBCT data obtained during RT contain considerable noise and subtle density changes in the low-dose region will be close to impossible to detect with the current image quality. Therefore, the region of lung receiving less than 20 Gy was removed for the extraction of the CBCT markers because it potentially could dilute the CBCT response signal. The upper dose level of 45 Gy was chosen to match the region of lung analysed in the follow-up CT dataset. To quantify the response, CBCBMthreshold was calculated as follows. The patient specific noise level (NL) was determined as NL ¼ 1:96 SD; where SD is the standard deviation of the CBCT density change observed at the 2nd fraction. Hence, density changes observed this early in the treatment course are attributed to noise alone. At a subsequent fraction, a threshold value, T, was defined as, T ¼ P þ NL, where P is the peak position of the histogram at that fraction. The peak position is included to adjust for daily variations of the CBCT image intensity. CBCTMthreshold was defined as the fraction of voxels in the histogram above this threshold (see Supplementary Fig. S1b). Statistical analysis Univariable correlations with LuRa were calculated using Spearman’s rank. Parameters investigated were: CBCTMslope, CBCTMthreshold, gender, age, performance status >1, forced expiratory volume in 1 s (FEV1) prior to radiotherapy, smoking status, prescribed dose, mean lung dose, omission of chemotherapy, time to follow-up, and baseline lung density. A linear regression model was used where CT evaluated lung responsiveness LuRa was log transformed to reach homogeneity of variance according to
y¼
jLuRaj logðjLuRaj þ 1Þ: LuRa
ð1Þ
The log-modulus transformation in Eq. (1) was used in place of a simple log transformation as some of the LuRa values are negative [11]. In a multivariable analysis, the linear regression included, in turn, one of the CBCTM biomarkers as well as those of the parameters specified above that reaches 10% significance in the univariable analysis. Stepwise forward selection and backward elimination was performed by use of the likelihood ratio method as evaluation. Entry and removal value was 5%. The 85th percentile of the LuRa values was used to dichotomize patients as low- or high-responding. This value was chosen to mimic the rate of severe RP in NSCLC patients treated with intensity modulated RT [12]. The ability of the models to classify patients as low-responsive correctly was analysed by receiver operating characteristic (ROC) curves. DeLong’s test was used for difference in area under the curve (AUC) [13]. McNemar’s test was used for the difference in number of correctly classified patients. Error bars represent 95% confidence intervals. Two-tailed p-values less than 0.05 were considered statistically significant.
positive LuRa value, but there is a tail to the right indicating a few highly responsive patients. The values range from 1.05 to 6.18 HU/Gy with a median at 0.97 HU/Gy. Univariable correlations between LuRa and investigated patient parameters are shown in Table 2. None of the CBCTMslope imaging biomarkers exhibited significant correlation with LuRa and CBCTM will in the following refer to CBCTMthreshold. CBCTM obtained at the 20th and 30th fraction were significantly correlated with LuRa while the marker obtained at the 10th fraction only reached borderline significance (p = 0.068). The correlation coefficients increase with increasing fraction number. Of the baseline clinical characteristics, age and omission of both concomitant and neoadjuvant chemotherapy appears positively correlated with LuRa, while current smoking was found to be negatively correlated with LuRa. The negative correlation between smoking status and LuRa indicates that current smoking reduces the sensitivity of the lung tissue. However, both smoking status and omission of chemotherapy were correlated with age, and these findings may therefore be spurious due to confounding effects. In panel (a) of Fig. 2, LuRa is plotted as function of CBCTM values obtained at the 10th, 20th, and 30th fraction. Age and CBCTM were significant in the multivariable analysis resulting in a model for the predicted lung radioresponsiveness (pLuRa) given by
pLuRa ¼
jyj jyj e 1 ; y
y ¼ k0 þ kage age þ kCBCTM CBCTM:
The maximal obtained LuRa for the 135 NSCLC patients are found in the histogram in Fig. 1. Most patients have a small
ð3Þ
The estimated coefficient values are shown in Table 3. Both stepwise forward selection and backward elimination resulted in the selection of these variables. As seen in the table, Spearman’s rank correlation coefficient for the models was 0.35, 0.35, and 0.39, when based on the 10th, 20th, and 30th fraction CBCTM values, respectively, and the effect size dominated by age. A graphical presentation of the model is shown in Fig. 2 panel (b), where observed LuRa is plotted against pLuRa. Dichotomization at the 85th percentile of LuRa corresponded to 2.7 HU/Gy. Thus, patients with LuRa values in the tail to the right in the histogram in Fig. 1 were assigned as high-responsive. ROC curves for classifying patients as low-responsive correctly are shown in Fig. 3. The AUC for the full models was not significantly larger than the AUC using age alone as predictor (p = 0.17–0.26) as shown in Supplementary Table 1. However, in the regime where all high-responsive patients are classified correctly (specificity = 1), the full model classified significantly more low-responsive patients correctly than the model only including age. Age alone classified 36% of the low-responsive patients correctly while the 10th
60 50 40 30 20 10 0
Results
ð2Þ
with
Number of patients
of a linear fit to the dose–effect curve determined by the CBCT density changes at a given fraction. CBCTMthreshold was extracted by histogram analysis of the density changes as described briefly below and in detail in the Supplementary material.
−2
0
2 4 LuRa [HU/Gy]
6
8
Fig. 1. Histogram of the maximal obtained lung radioresponsiveness (LuRa) values for the 135 patients.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
4
Predicting lung density changes after RT
Table 2 Univariable correlations with lung radioresponsiveness. Parameter
Spearman’s rank
Gender (Male) Age Performance Status > 1 FEV1 Smoking Total radiation dose Mean lung dose Omission of chemotherapy Only neoadjuvant chemotherapy Time to follow-up Baseline CT density CBCTMslope at 10th fraction CBCTMslope at 20th fraction CBCTMslope at 30th fraction CBCTMthreshold at 10th fraction CBCTMthreshold at 20th fraction CBCTMthreshold at 30th fraction
R
p-Value
0.002 0.310 0.057 0.038 0.210 0.000 0.002 0.179 0.085 0.089 0.137 0.034 0.036 0.074 0.159 0.185 0.339
0.979 <0.001 0.513 0.666 0.015 0.996 0.978 0.038 0.327 0.306 0.114 0.695 0.683 0.432 0.068 0.036 <0.001
Abbreviations: FEV1 = forced expiratory volume in 1 s. Parameters significantly correlated to lung radioresponsiveness at the 5% level are shown in bold.
fraction model was most successful with 58% of the low-responsive patients classified correctly (see Supplementary Table 1).
Discussion The data presented in this study suggest that it is possible to predict the LuRa of individual patients using age and CBCTM values obtained during treatment, but the correlations of predicted versus observed LuRa are low which indicates that the models might have limited clinical impact. As seen in Table 3, age had a larger
LuRa [HU/Gy]
(a)
10th fraction
30th fraction
20th fraction
6
6
6
4
4
4
2
2
2
0
0
0
-2
0
(b) LuRa [HU/Gy]
(although not statistically significant) influence on pLuRa than CBCTM. The ROC analysis was performed to investigate whether the response markers could have a significant impact on identifying low-responsive patients. The results indicate that adding CBCTM does not increase the AUC of the model compared to age alone. However, for safe dose escalation most high-responsive patients must be identified correctly. This corresponds to the high-specificity region of the ROC curve. When the specificity is 1, the full model including CBCTM classified more than half of the low-responsive patients correctly already at the 10th and 20th treatment fraction, while age alone only classified 36% correctly. Perhaps a bit surprising, the model based on imaging already at the 10th fraction classified 58% of the low-responsive patients correctly. Hence, in an adaptive RT scenario, the effect of adding the response markers in the model might have clinical impact. The end-point considered in the present study was the slope of a linear fit to the dose–effect curves relating changes in regional lung density observed at follow-up to the regional planned dose. LuRa is thus a measure of the patients’ individual susceptibility to radiation in terms of density changes. The continuous nature of this end-point makes it ideal for radiobiological response modelling. However, the radiological changes may be sub-clinical, and although RP has been shown to correlate with the development of radiologic consolidation on CT [4], a firm association between clinically scored RP and CT density changes has yet to be established for lung cancer patients [5,14]. Low-responsive patients are therefore not necessarily free from symptomatic RP and vice versa. The patients of the current study have not been scored for RP and the presented models needs validation in a cohort of NSCLC patients with recorded toxicity. The relation between LuRa and age in our study is not surprising. Ma et al. [15] found that greater age and pre-RT surgery
0.1
0.2
0.3
-2
0
0.1 0.2 CBCTM
0.3
-2
6
6
6
4
4
4
2
2
2
0
0
0
0
1
2
3
4
-2
0
1 2 3 pLuRa [HU/Gy]
0.1
0.2
0.3
30th fraction model
20th fraction model
10th fraction model
-2
0
4
-2
0
1
2
3
4
Fig. 2. (a) Lung radioresponsiveness (LuRa) plotted as function of cone-beam computed tomography markers (CBCTM) values obtained at the 10th, 20th, and 30th fraction. (b) LuRa plotted against predicted lung radioresponsiveness (pLuRa). The pLuRa values were binned in the quartiles of LuRa values, and the median in each bin is shown in the plot. The identity is shown as the dashed line.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
5
U. Bernchou et al. / Radiotherapy and Oncology xxx (2015) xxx–xxx Table 3 Multivariable regression analysis. jyj pLuRa ¼ jyj 1Þ; y ¼ k0 þ kage age þ kCBCTM CBCTM y ðe
Parameter
Value
P
10th fraction model k0 kage kCBCTM
0.80 (1.48;0.13) 0.021 (0.011;0.031) 3.53 (0.27;6.80)
0.022 <0.001 0.036
20th fraction model k0 kage kCBCTM
0.76 (1.43;0.09) 0.020 (0.010;0.030) 3.46 (0.41;6.50)
0.028 <0.001 0.028
30th fraction model k0 kage kCBCTM
0.74 (1.40;0.09) 0.019 (0.010;0.029) 3.12 (0.074;5.10)
0.028 <0.001 0.003
Effect size
Spearman’s rank R
p
0.35
<0.001
0.35
<0.001
0.39
<0.001
0.42 (0.11;1.26) 0.20 (0.01;0.45)
0.40 (0.10;1.22) 0.21 (0.02;0.47)
0.38 (0.09;1.17) 0.30 (0.09;0.57)
Predicted lung radioresponsiveness (pluRa) has units of [HU/Gy]. The effect size is defined as the change in pLuRa when changing a parameter from the median value to the median value plus one standard deviation of the parameter, while keeping all other parameters fixed and has the units of [HU/Gy]. kage has units of [year1]. The other parameters are unitless. The standard deviation of the age is 8.6 years and the standard deviation of the CBCTM values is 0.026, 0.028, and 0.042 for 10th, 20th, and 30th fraction, respectively.
1.0
Sensitivity
0.8
0.6
0.4
Age 10th fraction model 20th fraction model 30th fraction model Reference Line
0.2
AUC 0.77 0.82 0.82 0.81 0.5
0.0 0.0
0.2
0.4
0.6
0.8
1.0
1 - Specificity Fig. 3. Receiver operating characteristic (ROC) curves for classifying patients as low-responsive. The area under the curve (AUC) for the models is shown in the figure.
increased the lung density in the irradiated volume in a cohort of lung, breast, and lymphoma patients in a multivariable analysis. To test whether co-morbidity is confounding the effect of age in our model, we investigated the relation between the baseline lung density and LuRa, as patients with chronic obstructive lung disease and pulmonary emphysema generally have lower lung density. We did not see a significant univariable relation between LuRa and baseline density although the trend is in the expected direction (Spearman’s R = 0.14, p = 0.114). Furthermore, Tsujino et al. [16] recently found age as an independent risk factor for developing RP on a multivariable analysis in a study that also included pulmonary fibrosis score based on honeycombing and pulmonary emphysema score based on the extent of the low density areas in the peripheral lung on baseline CT.
The CBCT images used in the present study were acquired routinely for patient positioning during the fractionated treatment course and using a fast gantry speed [17]. The extraction of the markers does not require extra labour for treatment staff or additional imaging dose for the patient. The image quality may be increased by use of Monte Carlo based scatter correction [18] or by slowing the gantry speed during imaging which, potentially, could lead to improved predictive ability of the markers. Furthermore, the present results could potentially be combined with texture measures as proposed by Cunliffe et al. [19,20]. To summarize, this study indicates that response markers extracted from CBCT images acquired for IGRT have the potential to improve the prediction of LuRa in NSCLC patients. Age and response markers extracted at 10th, 20th, and 30th treatment
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
6
Predicting lung density changes after RT
fraction were included in a set of response models for predicting LuRa. The models suggest that younger patients without lung tissue reactions early into their treatment course may have minimal density increase during the first 6 months after commencement of RT. The proposed models need validation in a cohort where symptomatic toxicity data have been collected to be used for identifying low-responsive patients who are potential candidates for safe dose escalation. Conflicts of interest statement None. Acknowledgements This work was supported by The Lundbeck Foundation Center for Interventional Research in Radiation Oncology (CIRRO), The Danish Council for Strategic Research, and AgeCare (Academy of Geriatric Cancer Research), an international research collaboration based at Odense University Hospital, Denmark. UB acknowledges funding from Elekta. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.radonc.2015.07. 021. References [1] Marks LB, Bentzen SM, Deasy JO, et al. Radiation dose-volume effects in the lung. Int J Radiat Oncol Biol Phys 2010;76:S70–6. [2] Bentzen SM, Parliament M, Deasy JO, et al. Biomarkers and surrogate endpoints for normal-tissue effects of radiation therapy: the importance of dose-volume effects. Int J Radiat Oncol Biol Phys 2010;76:S145–50. [3] Bertelsen A, Schytte T, Bentzen SM, Hansen O, Nielsen M, Brink C. Radiation dose response of normal lung assessed by Cone Beam CT – a potential tool for biologically adaptive radiation therapy. Radiother Oncol 2011;100:351–5. [4] Jenkins P, Welsh A. Computed tomography appearance of early radiation injury to the lung: correlation with clinical and dosimetric factors. Int J Radiat Oncol Biol Phys 2011;81:97–103.
[5] De Ruysscher D, Sharifi H, Defraene G, et al. Quantification of radiationinduced lung damage with CT scans: the possible benefit for radiogenomics. Acta Oncol 2013;52:1405–10. [6] Bernchou U, Schytte T, Bertelsen A, Bentzen SM, Hansen O, Brink C. Time evolution of regional CT density changes in normal lung after IMRT for NSCLC. Radiother Oncol 2013;109:89–94. [7] Feldkamp L, Davis L, Kress J. Practical cone-beam algorithm. J Opt Soc Am A: 1984;1:612–9. [8] Rit S, Vila Oliva M, Brousmiche S, Labarbe R, Sarrut D, and Sharp GC. The Reconstruction Toolkit (RTK), an open-source cone-beam CT reconstruction toolkit based on the Insight Toolkit (ITK). Journal of Physics: Conference Series, 489:012079, 2014. [9] Brink C, Bernchou U, Bertelsen A, Hansen O, Schytte T, Bentzen SM. Locoregional control of non-small cell lung cancer in relation to automated early assessment of tumor regression on cone beam computed tomography. Int J Radiat Oncol Biol Phys 2014;89:916–23. [10] Klein S, Staring M, Murphy K, Viergever MA, Pluim JPW. Elastix: a toolbox for intensity-based medical image registration. IEEE Trans Med Imaging 2010;29:196–205. [11] John JA, Draper NR. An alternative family of transformations. Appl. Statist. 1980;29:190–7. [12] Jiang ZQ, Yang K, Komaki R, et al. Long-term clinical outcome of intensitymodulated radiotherapy for inoperable non-small cell lung cancer: the MD Anderson experience. Int J Radiat Oncol Biol Phys 2012;83:332–9. [13] DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44:837–45. [14] Phernambucq ECJ, Palma DA, Vincent A, Smit EF, Senan S. Time and doserelated changes in radiological lung density after concurrent chemoradiotherapy for lung cancer. Lung cancer 2011;74:451–6. [15] Ma J, Zhang J, Zhou S, et al. Regional lung density changes after radiation therapy for tumors in and around thorax. Int J Radiat Oncol Biol Phys 2010;76:116–22. [16] Tsujino K, Hashimoto T, Shimada T, et al. Combined analysis of V20, VS5, pulmonary fibrosis score on baseline computed tomography, and patient age improves prediction of severe radiation pneumonitis after concurrent chemoradiotherapy for locally advanced non-small-cell lung cancer. J Thorac Oncol 2014;9:983–90. [17] Westberg J, Jensen HR, Bertelsen A, Brink C. Reduction of Cone-Beam CT scan time without compromising the accuracy of the image registration in IGRT. Acta Oncol 2010;49:225–9. [18] Thing RS, Bernchou U, Mainegra-Hing E, Brink C. Patient-specific scatter correction in clinical cone beam computed tomography imaging made possible by the combination of Monte Carlo simulations and a ray tracing algorithm. Acta Oncol 2013;52:1477–83. [19] Cunliffe AR, Armato 3rd SG, Straus C, Malik R, Al-Hallaq HA. Lung texture in serial thoracic CT scans: correlation with radiologist-defined severity of acute changes following radiation therapy. Phys Med Biol 2014;59:5387–98. [20] Cunliffe A, Armato 3rd SG, Castillo R, Pham N, Guerrero T, Al-Hallaq HA. Lung texture in serial thoracic computed tomography scans: Correlation of radiomics-based features with radiation therapy dose and radiation pneumonitis development. Int J Radiat Oncol Biol Phys 2015.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
Contents lists available at ScienceDirect
Radiotherapy and Oncology journal homepage: www.thegreenjournal.com
Original article
Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients Uffe Bernchou a,b,⇑, Olfred Hansen a,c, Tine Schytte c, Anders Bertelsen b, Andrew Hope d, Douglas Moseley d, Carsten Brink a,b a d
Institute of Clinical Research, University of Southern Denmark, Odense; b Laboratory of Radiation Physics; Department of Radiation Oncology, Princess Margaret Hospital, Toronto, Canada
a r t i c l e
i n f o
Article history: Received 11 May 2015 Received in revised form 30 June 2015 Accepted 16 July 2015 Available online xxxx Keywords: Lung cancer Density change Computed tomography Cone beam Response modelling
c
Department of Oncology, Odense University Hospital, Denmark ; and
a b s t r a c t Background and purpose: This study investigates the ability of pre-treatment factors and response markers extracted from standard cone-beam computed tomography (CBCT) images to predict the lung density changes induced by radiotherapy for non-small cell lung cancer (NSCLC) patients. Methods and materials: Density changes in follow-up computed tomography scans were evaluated for 135 NSCLC patients treated with radiotherapy. Early response markers were obtained by analysing changes in lung density in CBCT images acquired during the treatment course. The ability of pre-treatment factors and CBCT markers to predict lung density changes induced by radiotherapy was investigated. Results: Age and CBCT markers extracted at 10th, 20th, and 30th treatment fraction significantly predicted lung density changes in a multivariable analysis, and a set of response models based on these parameters were established. The correlation coefficient for the models was 0.35, 0.35, and 0.39, when based on the markers obtained at the 10th, 20th, and 30th fraction, respectively. Conclusions: The study indicates that younger patients without lung tissue reactions early into their treatment course may have minimal radiation induced lung density increase at follow-up. Further investigations are needed to examine the ability of the models to identify patients with low risk of symptomatic toxicity. Ó 2015 Elsevier Ireland Ltd. All rights reserved. Radiotherapy and Oncology xxx (2015) xxx–xxx
Dose escalation for non-small cell lung cancer (NSCLC) is typically limited by increasing risk of severe adverse events such as radiation pneumonitis (RP). Therefore, identifying patients with low or high risk of developing normal tissue toxicity is of importance if the dose is to be escalated safely. Response predictions are typically based on models including dosimetric parameters [1]. However, the predictive power of the models could potentially be increased by adding other baseline clinical factors and predictive markers obtained prior to treatment or response markers acquired early into the treatment course [2]. A recent study has detected normal tissue reactions in terms of density change in healthy lung tissue in cone-beam computed tomography (CBCT) images obtained relatively early during a fractionated radiotherapy (RT) treatment course [3]. Although lung density change not necessarily implies symptomatic toxicity, RP ⇑ Corresponding author at: Laboratory of Radiation Physics, Odense University Hospital, Sdr. Boulevard 29, DK-5000 Odense C, Denmark. E-mail address: [email protected] (U. Bernchou).
typically manifests as increases in density within the healthy lung tissue as observed in follow-up computed tomography (CT) images [4]. In the current study, density changes in follow-up CT were evaluated in a cohort of NSCLC patients treated with CBCT image guided RT. Lung Radioresponsiveness (LuRa) was defined as the slope of a linear fit to dose–effect curves produced by relating changes in regional lung density to regional planned dose as proposed by De Ruysscher et al. [5]. LuRa thus provides an objective, continuous measure of the patients’ individual susceptibility to radiation in terms of density change. The current study investigated the ability of pre-treatment factors and CBCT markers (CBCTM) to predict LuRa. Methods and materials The study includes NSCLC patients treated to a prescribed dose of 60 or 66 Gy in 2-Gy fractions in the period from 2007–2013 at Odense University Hospital. A total of 135 patients (Table 1) met the following inclusion criteria: (1) CBCTs were acquired
http://dx.doi.org/10.1016/j.radonc.2015.07.021 0167-8140/Ó 2015 Elsevier Ireland Ltd. All rights reserved.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
2
Predicting lung density changes after RT
Table 1 Characteristics of the 135 patients. Characteristic
na
%
Gender Male Female
71 64
53 47
Age (years) Median Range
68 33–85
Performance (WHO) 0 1 2 3
40 70 24 1
FEV1 (l/min) Median Range
1.8 0.6–4.5
Stage 1–2 3 4 Recurrence
27 92 5 11
20 68 4 8
Histology Squamous cell carcinoma Adenocarcinoma Otherb
63 54 18
47 40 13
Current smoking status Smoking Not smoking
71 64
53 47
GTV volume (cm3) Median Range
31 0–400
PTV volume (cm3) Median Range Total radiation dose (Gy) 60 66
29 52 18 1
Image registration
385 62–1648 36 99
Mean lung dose (Gy) Median Range
17 2–24
Chemotherapy No Only neoadjuvant Only concomitant Both neoadjuvant and concomitant
30 18 1 86
groups of treatment fractions were analysed: (I) 1, (II) 2–3, (III) 10–13, (IV) 20–23, and (V) 30–33. To be included in the study a patient must have had at least one CBCT acquired in groups (I) and (II) and at least one CBCT in either of the groups (III)–(V), and all CBCTs must be acquired on the same accelerator and using the same imaging technique and filter settings. If several scans were acquired in each group, the earliest scan was used. As described below, density changes in CBCT images were monitored relative to the CBCT density at the 1st fraction, while the scans in group (II) were included to measure the patient specific noise level, and response markers were extracted from the scans in group (III)– (V). The fraction numbers within each group were chosen to match the action limit protocol with a few extra scans included to increase the number of valid scans in the cases where the imaging protocol or the accelerator was changed. Scans belonging to group (II)–(V) will in the following be termed 2nd, 10th, 20th, and 30th fraction, respectively. The number of CBCT scans available was 133, 129, and 113 at the 10th, 20th, and 30th fraction, respectively, and in total 645 CBCT scans were analysed. The clinical scans were either 3D or 4D CBCT. However, for each scan used in the study, the raw projection images were reconstructed to 3D with a slice thickness of 2 mm and pixel size of 1 mm by the Feldkamp–David–Kress [7] algorithm implemented in the open-source cone beam Reconstruction Toolkit [8]. A total of 222 follow-up CT scans acquired within 6 months from commencement of RT were evaluable for the LuRa determination (see details below). The follow-up CT scans were typically acquired in the inspirational breath-hold at a diagnostic CT scanner using a median slice thickness of 5 mm (range 2–7.5 mm) and a pixel size of 0.7 mm (range 0.6–1.0 mm).
27 73
22 13 1 64
Abbreviations: GTV = gross tumour volume; PTV = planned target volume. a n denoted number of patients unless otherwise indicated. b Large cell carcinoma, undifferentiated, atypical carcinoid, unspecified.
throughout the treatment as specified below. (2) Eligible follow-up CT scans were acquired within 6 months from commencement of RT without clinical or radiological evidence of local recurrence or thoracic metastases. (3) Acceptable image registration, as described below, was evaluated on visual validation. (4) The volume of investigated healthy lung tissue was larger than 250 cm3, to ensure sufficient data for the individual patient. (5) No sign of atelectasis was seen in the investigated lung tissue volume during treatment or at follow-up. Details on treatment planning, chemotherapy and follow-up are described elsewhere [6]. Imaging data Treatment planning was based on the mid-ventilation phase of a 4D CT scan with slice thickness of 2.5 or 3 mm and pixel size of 1 mm. RT was delivered on Elekta Synergy accelerators equipped with an X-ray Volume Imaging (XVI) system. CBCT image guided RT (IGRT) was performed either daily or at the first three fractions and at least at the 10th, 20th, and 30th fraction, according to an action limit protocol. Imaging data acquired at the following five
The main steps in assessing density changes in healthy lung tissue from CBCT and follow-up CT image data and relating this to the locally delivered dose have been described in detail elsewhere [3,6,9]. Briefly, deformable image registrations were performed using the elastix freeware toolbox [10]. CBCT-CBCT registrations allow the creation of 3D maps of CBCT density changes throughout the treatment by subtraction of the images acquired at the 1st fraction from the images acquired at subsequent fractions. By registration of the CBCTs to the planning CT, the density changes were related to the locally planned dose. Similarly, the follow-up CT scans were co-registered to the planning CT for the investigation of lung density changes post treatment. For both follow-up CT and CBCT, a 5 5 5 mm wide median filter was applied on the 3D maps of density change in order to suppress noise. Data analysis Density changes were evaluated in the regions of healthy lung tissue within the CBCT field of view. Healthy lung tissue was defined as the total lung delineated during treatment planning, eroded by 10 mm and with the PTV subtracted to avoid boundary effects. Patient individual dose–effect curves at follow-up were created by averaging the density changes post treatment for regions receiving the same planned dose in 5-Gy bins. The LuRa for a patient was defined as the slope of a linear regression to the curve as suggested by De Ruysscher et al. [5]. The high dose regions (>45 Gy) of the curve were excluded from the fit as previous investigations have shown that dose response is non-linear in this regime, at least in the early regime (<3 months) [6]. For patients with more than one evaluable follow-up scan, the highest LuRa value was chosen. Two types of response markers were extracted from the cone beam images: a slope marker (CBCTMslope) and a threshold marker (CBCTMthreshold). CBCTMslope was calculated similar to the LuRa values at follow-up as the slope
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
3
U. Bernchou et al. / Radiotherapy and Oncology xxx (2015) xxx–xxx
Histogram analysis Voxel histograms of the density change relative to the CBCT acquired at the 1st fraction in regions receiving 20–45 Gy planned dose were produced for each available treatment fraction (2nd-30th) for each patient. Although density changes have been observed in the low-dose region in follow-up CT scans [6], the CBCT data obtained during RT contain considerable noise and subtle density changes in the low-dose region will be close to impossible to detect with the current image quality. Therefore, the region of lung receiving less than 20 Gy was removed for the extraction of the CBCT markers because it potentially could dilute the CBCT response signal. The upper dose level of 45 Gy was chosen to match the region of lung analysed in the follow-up CT dataset. To quantify the response, CBCBMthreshold was calculated as follows. The patient specific noise level (NL) was determined as NL ¼ 1:96 SD; where SD is the standard deviation of the CBCT density change observed at the 2nd fraction. Hence, density changes observed this early in the treatment course are attributed to noise alone. At a subsequent fraction, a threshold value, T, was defined as, T ¼ P þ NL, where P is the peak position of the histogram at that fraction. The peak position is included to adjust for daily variations of the CBCT image intensity. CBCTMthreshold was defined as the fraction of voxels in the histogram above this threshold (see Supplementary Fig. S1b). Statistical analysis Univariable correlations with LuRa were calculated using Spearman’s rank. Parameters investigated were: CBCTMslope, CBCTMthreshold, gender, age, performance status >1, forced expiratory volume in 1 s (FEV1) prior to radiotherapy, smoking status, prescribed dose, mean lung dose, omission of chemotherapy, time to follow-up, and baseline lung density. A linear regression model was used where CT evaluated lung responsiveness LuRa was log transformed to reach homogeneity of variance according to
y¼
jLuRaj logðjLuRaj þ 1Þ: LuRa
ð1Þ
The log-modulus transformation in Eq. (1) was used in place of a simple log transformation as some of the LuRa values are negative [11]. In a multivariable analysis, the linear regression included, in turn, one of the CBCTM biomarkers as well as those of the parameters specified above that reaches 10% significance in the univariable analysis. Stepwise forward selection and backward elimination was performed by use of the likelihood ratio method as evaluation. Entry and removal value was 5%. The 85th percentile of the LuRa values was used to dichotomize patients as low- or high-responding. This value was chosen to mimic the rate of severe RP in NSCLC patients treated with intensity modulated RT [12]. The ability of the models to classify patients as low-responsive correctly was analysed by receiver operating characteristic (ROC) curves. DeLong’s test was used for difference in area under the curve (AUC) [13]. McNemar’s test was used for the difference in number of correctly classified patients. Error bars represent 95% confidence intervals. Two-tailed p-values less than 0.05 were considered statistically significant.
positive LuRa value, but there is a tail to the right indicating a few highly responsive patients. The values range from 1.05 to 6.18 HU/Gy with a median at 0.97 HU/Gy. Univariable correlations between LuRa and investigated patient parameters are shown in Table 2. None of the CBCTMslope imaging biomarkers exhibited significant correlation with LuRa and CBCTM will in the following refer to CBCTMthreshold. CBCTM obtained at the 20th and 30th fraction were significantly correlated with LuRa while the marker obtained at the 10th fraction only reached borderline significance (p = 0.068). The correlation coefficients increase with increasing fraction number. Of the baseline clinical characteristics, age and omission of both concomitant and neoadjuvant chemotherapy appears positively correlated with LuRa, while current smoking was found to be negatively correlated with LuRa. The negative correlation between smoking status and LuRa indicates that current smoking reduces the sensitivity of the lung tissue. However, both smoking status and omission of chemotherapy were correlated with age, and these findings may therefore be spurious due to confounding effects. In panel (a) of Fig. 2, LuRa is plotted as function of CBCTM values obtained at the 10th, 20th, and 30th fraction. Age and CBCTM were significant in the multivariable analysis resulting in a model for the predicted lung radioresponsiveness (pLuRa) given by
pLuRa ¼
jyj jyj e 1 ; y
y ¼ k0 þ kage age þ kCBCTM CBCTM:
The maximal obtained LuRa for the 135 NSCLC patients are found in the histogram in Fig. 1. Most patients have a small
ð3Þ
The estimated coefficient values are shown in Table 3. Both stepwise forward selection and backward elimination resulted in the selection of these variables. As seen in the table, Spearman’s rank correlation coefficient for the models was 0.35, 0.35, and 0.39, when based on the 10th, 20th, and 30th fraction CBCTM values, respectively, and the effect size dominated by age. A graphical presentation of the model is shown in Fig. 2 panel (b), where observed LuRa is plotted against pLuRa. Dichotomization at the 85th percentile of LuRa corresponded to 2.7 HU/Gy. Thus, patients with LuRa values in the tail to the right in the histogram in Fig. 1 were assigned as high-responsive. ROC curves for classifying patients as low-responsive correctly are shown in Fig. 3. The AUC for the full models was not significantly larger than the AUC using age alone as predictor (p = 0.17–0.26) as shown in Supplementary Table 1. However, in the regime where all high-responsive patients are classified correctly (specificity = 1), the full model classified significantly more low-responsive patients correctly than the model only including age. Age alone classified 36% of the low-responsive patients correctly while the 10th
60 50 40 30 20 10 0
Results
ð2Þ
with
Number of patients
of a linear fit to the dose–effect curve determined by the CBCT density changes at a given fraction. CBCTMthreshold was extracted by histogram analysis of the density changes as described briefly below and in detail in the Supplementary material.
−2
0
2 4 LuRa [HU/Gy]
6
8
Fig. 1. Histogram of the maximal obtained lung radioresponsiveness (LuRa) values for the 135 patients.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
4
Predicting lung density changes after RT
Table 2 Univariable correlations with lung radioresponsiveness. Parameter
Spearman’s rank
Gender (Male) Age Performance Status > 1 FEV1 Smoking Total radiation dose Mean lung dose Omission of chemotherapy Only neoadjuvant chemotherapy Time to follow-up Baseline CT density CBCTMslope at 10th fraction CBCTMslope at 20th fraction CBCTMslope at 30th fraction CBCTMthreshold at 10th fraction CBCTMthreshold at 20th fraction CBCTMthreshold at 30th fraction
R
p-Value
0.002 0.310 0.057 0.038 0.210 0.000 0.002 0.179 0.085 0.089 0.137 0.034 0.036 0.074 0.159 0.185 0.339
0.979 <0.001 0.513 0.666 0.015 0.996 0.978 0.038 0.327 0.306 0.114 0.695 0.683 0.432 0.068 0.036 <0.001
Abbreviations: FEV1 = forced expiratory volume in 1 s. Parameters significantly correlated to lung radioresponsiveness at the 5% level are shown in bold.
fraction model was most successful with 58% of the low-responsive patients classified correctly (see Supplementary Table 1).
Discussion The data presented in this study suggest that it is possible to predict the LuRa of individual patients using age and CBCTM values obtained during treatment, but the correlations of predicted versus observed LuRa are low which indicates that the models might have limited clinical impact. As seen in Table 3, age had a larger
LuRa [HU/Gy]
(a)
10th fraction
30th fraction
20th fraction
6
6
6
4
4
4
2
2
2
0
0
0
-2
0
(b) LuRa [HU/Gy]
(although not statistically significant) influence on pLuRa than CBCTM. The ROC analysis was performed to investigate whether the response markers could have a significant impact on identifying low-responsive patients. The results indicate that adding CBCTM does not increase the AUC of the model compared to age alone. However, for safe dose escalation most high-responsive patients must be identified correctly. This corresponds to the high-specificity region of the ROC curve. When the specificity is 1, the full model including CBCTM classified more than half of the low-responsive patients correctly already at the 10th and 20th treatment fraction, while age alone only classified 36% correctly. Perhaps a bit surprising, the model based on imaging already at the 10th fraction classified 58% of the low-responsive patients correctly. Hence, in an adaptive RT scenario, the effect of adding the response markers in the model might have clinical impact. The end-point considered in the present study was the slope of a linear fit to the dose–effect curves relating changes in regional lung density observed at follow-up to the regional planned dose. LuRa is thus a measure of the patients’ individual susceptibility to radiation in terms of density changes. The continuous nature of this end-point makes it ideal for radiobiological response modelling. However, the radiological changes may be sub-clinical, and although RP has been shown to correlate with the development of radiologic consolidation on CT [4], a firm association between clinically scored RP and CT density changes has yet to be established for lung cancer patients [5,14]. Low-responsive patients are therefore not necessarily free from symptomatic RP and vice versa. The patients of the current study have not been scored for RP and the presented models needs validation in a cohort of NSCLC patients with recorded toxicity. The relation between LuRa and age in our study is not surprising. Ma et al. [15] found that greater age and pre-RT surgery
0.1
0.2
0.3
-2
0
0.1 0.2 CBCTM
0.3
-2
6
6
6
4
4
4
2
2
2
0
0
0
0
1
2
3
4
-2
0
1 2 3 pLuRa [HU/Gy]
0.1
0.2
0.3
30th fraction model
20th fraction model
10th fraction model
-2
0
4
-2
0
1
2
3
4
Fig. 2. (a) Lung radioresponsiveness (LuRa) plotted as function of cone-beam computed tomography markers (CBCTM) values obtained at the 10th, 20th, and 30th fraction. (b) LuRa plotted against predicted lung radioresponsiveness (pLuRa). The pLuRa values were binned in the quartiles of LuRa values, and the median in each bin is shown in the plot. The identity is shown as the dashed line.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
5
U. Bernchou et al. / Radiotherapy and Oncology xxx (2015) xxx–xxx Table 3 Multivariable regression analysis. jyj pLuRa ¼ jyj 1Þ; y ¼ k0 þ kage age þ kCBCTM CBCTM y ðe
Parameter
Value
P
10th fraction model k0 kage kCBCTM
0.80 (1.48;0.13) 0.021 (0.011;0.031) 3.53 (0.27;6.80)
0.022 <0.001 0.036
20th fraction model k0 kage kCBCTM
0.76 (1.43;0.09) 0.020 (0.010;0.030) 3.46 (0.41;6.50)
0.028 <0.001 0.028
30th fraction model k0 kage kCBCTM
0.74 (1.40;0.09) 0.019 (0.010;0.029) 3.12 (0.074;5.10)
0.028 <0.001 0.003
Effect size
Spearman’s rank R
p
0.35
<0.001
0.35
<0.001
0.39
<0.001
0.42 (0.11;1.26) 0.20 (0.01;0.45)
0.40 (0.10;1.22) 0.21 (0.02;0.47)
0.38 (0.09;1.17) 0.30 (0.09;0.57)
Predicted lung radioresponsiveness (pluRa) has units of [HU/Gy]. The effect size is defined as the change in pLuRa when changing a parameter from the median value to the median value plus one standard deviation of the parameter, while keeping all other parameters fixed and has the units of [HU/Gy]. kage has units of [year1]. The other parameters are unitless. The standard deviation of the age is 8.6 years and the standard deviation of the CBCTM values is 0.026, 0.028, and 0.042 for 10th, 20th, and 30th fraction, respectively.
1.0
Sensitivity
0.8
0.6
0.4
Age 10th fraction model 20th fraction model 30th fraction model Reference Line
0.2
AUC 0.77 0.82 0.82 0.81 0.5
0.0 0.0
0.2
0.4
0.6
0.8
1.0
1 - Specificity Fig. 3. Receiver operating characteristic (ROC) curves for classifying patients as low-responsive. The area under the curve (AUC) for the models is shown in the figure.
increased the lung density in the irradiated volume in a cohort of lung, breast, and lymphoma patients in a multivariable analysis. To test whether co-morbidity is confounding the effect of age in our model, we investigated the relation between the baseline lung density and LuRa, as patients with chronic obstructive lung disease and pulmonary emphysema generally have lower lung density. We did not see a significant univariable relation between LuRa and baseline density although the trend is in the expected direction (Spearman’s R = 0.14, p = 0.114). Furthermore, Tsujino et al. [16] recently found age as an independent risk factor for developing RP on a multivariable analysis in a study that also included pulmonary fibrosis score based on honeycombing and pulmonary emphysema score based on the extent of the low density areas in the peripheral lung on baseline CT.
The CBCT images used in the present study were acquired routinely for patient positioning during the fractionated treatment course and using a fast gantry speed [17]. The extraction of the markers does not require extra labour for treatment staff or additional imaging dose for the patient. The image quality may be increased by use of Monte Carlo based scatter correction [18] or by slowing the gantry speed during imaging which, potentially, could lead to improved predictive ability of the markers. Furthermore, the present results could potentially be combined with texture measures as proposed by Cunliffe et al. [19,20]. To summarize, this study indicates that response markers extracted from CBCT images acquired for IGRT have the potential to improve the prediction of LuRa in NSCLC patients. Age and response markers extracted at 10th, 20th, and 30th treatment
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021
6
Predicting lung density changes after RT
fraction were included in a set of response models for predicting LuRa. The models suggest that younger patients without lung tissue reactions early into their treatment course may have minimal density increase during the first 6 months after commencement of RT. The proposed models need validation in a cohort where symptomatic toxicity data have been collected to be used for identifying low-responsive patients who are potential candidates for safe dose escalation. Conflicts of interest statement None. Acknowledgements This work was supported by The Lundbeck Foundation Center for Interventional Research in Radiation Oncology (CIRRO), The Danish Council for Strategic Research, and AgeCare (Academy of Geriatric Cancer Research), an international research collaboration based at Odense University Hospital, Denmark. UB acknowledges funding from Elekta. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.radonc.2015.07. 021. References [1] Marks LB, Bentzen SM, Deasy JO, et al. Radiation dose-volume effects in the lung. Int J Radiat Oncol Biol Phys 2010;76:S70–6. [2] Bentzen SM, Parliament M, Deasy JO, et al. Biomarkers and surrogate endpoints for normal-tissue effects of radiation therapy: the importance of dose-volume effects. Int J Radiat Oncol Biol Phys 2010;76:S145–50. [3] Bertelsen A, Schytte T, Bentzen SM, Hansen O, Nielsen M, Brink C. Radiation dose response of normal lung assessed by Cone Beam CT – a potential tool for biologically adaptive radiation therapy. Radiother Oncol 2011;100:351–5. [4] Jenkins P, Welsh A. Computed tomography appearance of early radiation injury to the lung: correlation with clinical and dosimetric factors. Int J Radiat Oncol Biol Phys 2011;81:97–103.
[5] De Ruysscher D, Sharifi H, Defraene G, et al. Quantification of radiationinduced lung damage with CT scans: the possible benefit for radiogenomics. Acta Oncol 2013;52:1405–10. [6] Bernchou U, Schytte T, Bertelsen A, Bentzen SM, Hansen O, Brink C. Time evolution of regional CT density changes in normal lung after IMRT for NSCLC. Radiother Oncol 2013;109:89–94. [7] Feldkamp L, Davis L, Kress J. Practical cone-beam algorithm. J Opt Soc Am A: 1984;1:612–9. [8] Rit S, Vila Oliva M, Brousmiche S, Labarbe R, Sarrut D, and Sharp GC. The Reconstruction Toolkit (RTK), an open-source cone-beam CT reconstruction toolkit based on the Insight Toolkit (ITK). Journal of Physics: Conference Series, 489:012079, 2014. [9] Brink C, Bernchou U, Bertelsen A, Hansen O, Schytte T, Bentzen SM. Locoregional control of non-small cell lung cancer in relation to automated early assessment of tumor regression on cone beam computed tomography. Int J Radiat Oncol Biol Phys 2014;89:916–23. [10] Klein S, Staring M, Murphy K, Viergever MA, Pluim JPW. Elastix: a toolbox for intensity-based medical image registration. IEEE Trans Med Imaging 2010;29:196–205. [11] John JA, Draper NR. An alternative family of transformations. Appl. Statist. 1980;29:190–7. [12] Jiang ZQ, Yang K, Komaki R, et al. Long-term clinical outcome of intensitymodulated radiotherapy for inoperable non-small cell lung cancer: the MD Anderson experience. Int J Radiat Oncol Biol Phys 2012;83:332–9. [13] DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44:837–45. [14] Phernambucq ECJ, Palma DA, Vincent A, Smit EF, Senan S. Time and doserelated changes in radiological lung density after concurrent chemoradiotherapy for lung cancer. Lung cancer 2011;74:451–6. [15] Ma J, Zhang J, Zhou S, et al. Regional lung density changes after radiation therapy for tumors in and around thorax. Int J Radiat Oncol Biol Phys 2010;76:116–22. [16] Tsujino K, Hashimoto T, Shimada T, et al. Combined analysis of V20, VS5, pulmonary fibrosis score on baseline computed tomography, and patient age improves prediction of severe radiation pneumonitis after concurrent chemoradiotherapy for locally advanced non-small-cell lung cancer. J Thorac Oncol 2014;9:983–90. [17] Westberg J, Jensen HR, Bertelsen A, Brink C. Reduction of Cone-Beam CT scan time without compromising the accuracy of the image registration in IGRT. Acta Oncol 2010;49:225–9. [18] Thing RS, Bernchou U, Mainegra-Hing E, Brink C. Patient-specific scatter correction in clinical cone beam computed tomography imaging made possible by the combination of Monte Carlo simulations and a ray tracing algorithm. Acta Oncol 2013;52:1477–83. [19] Cunliffe AR, Armato 3rd SG, Straus C, Malik R, Al-Hallaq HA. Lung texture in serial thoracic CT scans: correlation with radiologist-defined severity of acute changes following radiation therapy. Phys Med Biol 2014;59:5387–98. [20] Cunliffe A, Armato 3rd SG, Castillo R, Pham N, Guerrero T, Al-Hallaq HA. Lung texture in serial thoracic computed tomography scans: Correlation of radiomics-based features with radiation therapy dose and radiation pneumonitis development. Int J Radiat Oncol Biol Phys 2015.
Please cite this article in press as: Bernchou U et al. Prediction of lung density changes after radiotherapy by cone beam computed tomography response markers and pre-treatment factors for non-small cell lung cancer patients. Radiother Oncol (2015), http://dx.doi.org/10.1016/j.radonc.2015.07.021