- Email: [email protected]
Adaptive Breast Radiation Therapy Using Modeling of Tissue Mechanics: A Breast Tissue Segmentation Study
International Journal of
Radiation Oncology biology
physics
www.redjournal.org
Physics Contribution
Adaptive Breast Radiation Therapy Using Modeling of Tissue Mechanics: A Breast Tissue Segmentation Study Prabhjot Juneja, MSc,* Emma J. Harris, PhD,* Anna M. Kirby, FRCR,y and Philip M. Evans, DPhil* *Joint Department of Physics, Institute of Cancer Research, Sutton, United Kingdom; and yDepartment of Academic Radiotherapy, Royal Marsden National Health Service Foundation Trust, Sutton, United Kingdom Received Dec 8, 2011, and in revised form Apr 26, 2012. Accepted for publication May 6, 2012
Summary Modeling of tissue mechanics coupled with CBCT may be used to predict the requirement for replanning (ie, adaptive radiation therapy). To model patients’ breasts, component tissues need to be segmented and assigned material properties. This study validates tissue segmentation methods and evaluates the effect of tissue distribution on segmentation accuracy. We found the most accurate method to be fuzzy c-means clustering (3 classes) and that the segmentation accuracy is related to the distribution of fibroglandular tissue.
Purpose: To validate and compare the accuracy of breast tissue segmentation methods applied to computed tomography (CT) scans used for radiation therapy planning and to study the effect of tissue distribution on the segmentation accuracy for the purpose of developing models for use in adaptive breast radiation therapy. Methods and Materials: Twenty-four patients receiving postlumpectomy radiation therapy for breast cancer underwent CT imaging in prone and supine positions. The whole-breast clinical target volume was outlined. Clinical target volumes were segmented into fibroglandular and fatty tissue using the following algorithms: physical density thresholding; interactive thresholding; fuzzy c-means with 3 classes (FCM3) and 4 classes (FCM4); and k-means. The segmentation algorithms were evaluated in 2 stages: first, an approach based on the assumption that the breast composition should be the same in both prone and supine position; and second, comparison of segmentation with tissue outlines from 3 experts using the Dice similarity coefficient (DSC). Breast datasets were grouped into nonsparse and sparse fibroglandular tissue distributions according to expert assessment and used to assess the accuracy of the segmentation methods and the agreement between experts. Results: Prone and supine breast composition analysis showed differences between the methods. Validation against expert outlines found significant differences (P<.001) between FCM3 and FCM4. Fuzzy c-means with 3 classes generated segmentation results (mean DSC Z 0.70) closest to the experts’ outlines. There was good agreement (mean DSC Z 0.85) among experts for breast tissue outlining. Segmentation accuracy and expert agreement was significantly higher (P<.005) in the nonsparse group than in the sparse group. Conclusions: The FCM3 gave the most accurate segmentation of breast tissues on CT data and could therefore be used in adaptive radiation therapy-based on tissue modeling. Breast tissue segmentation methods should be used with caution in patients with sparse fibroglandular tissue distribution. Ó 2012 Elsevier Inc.
Reprint requests to: Prabhjot Juneja, MSc, Institute of Cancer Research, Joint Department of Physics, Sutton SM2 5PT, United Kingdom. Tel: (þ44) 208-661-3490; Fax: (þ44) 208-643-3812; E-mail: Prabhjot. [email protected] P.J. is supported by an Institute of Cancer Research PhD studentship. This work was supported by Cancer Research-UK under program grant reference no. C46/A2131. Int J Radiation Oncol Biol Phys, Vol. 84, No. 3, pp. e419ee425, 2012 0360-3016/$ - see front matter Ó 2012 Elsevier Inc. All rights reserved. doi:10.1016/j.ijrobp.2012.05.014
Conflict of interest: none. Supplementary material for this article can be found at www.redjournal.org. AcknowledgmentsdThe authors thank Galvin Polundniowski and Imogen Locke for valuable discussions, and Gigin Lin and Tarun Durga for tissue outlining.
e420 Juneja et al.
International Journal of Radiation Oncology Biology Physics
Introduction Adjuvant breast radiation therapy after breast-conserving surgery in early-stage breast cancer patients has been shown to reduce the risk of local recurrence (1). The region proximal to the excision cavity, the “tumor bed” (TB), is the most likely place for recurrence and often receives escalated dose. Many imaging studies have shown that the TB volume changes significantly over the time-frame relevant to radiation therapy planning and delivery (2, 3). Adaptive radiation therapy (ART) is used to correct for changes in the TB and involves repeat computed tomography (CT) and replanning of the treatment. The workload related to reimaging and replanning is high. Currently image guided breast radiation therapy uses portal imaging or cone beam CT (CBCT) to visualize bony anatomy or fiducial markers as surrogates for the TB. These techniques are inadequate for detecting the actual changes in TB and other breast tissues. In this work we use modeling of tissue changes based on the finite element method (FEM) to model tissue changes and information from the planning CT (higher quality than CBCT), to understand how the breast tissues deform and move. At each treatment fraction the planning CT is deformably registered to the CBCT using the FEM model to physically constrain the deformable registration. This allows more accurate quantification of TB changes and their effect on the treatment, thereby determining the requirement for replanning. This provides a means for adaptive breast radiation therapy. The proposed ART workflow is shown in Fig. 1 as 9 steps. (1) The patient has a planning CT. (2) An expert outlines whole breast (WB), TB, seroma, and scar on the images from step 1. (3) Fibroglandular and fatty tissues are segmented in the images using algorithms developed in this work. (4) Tissues are assigned mechanical properties, and a patient-specific FEM model is constructed from step 3. CT numbers of scar and seroma are similar to fibroglandular tissue and therefore are segmented together. The expert outlines of scar and seroma are used to separate them from fibroglandular tissues, and the deformation of these tissues can be deduced without the need for an additional planning CT scan. (5) The treatment starts, and at each treatment fraction the patient has CBCT. (6) The FEM model is used to register planning CT (step 1) and CBCT (step 5). (7) The deformation in step 6 is used to determine the TB deformation. (8) Tissue deformations in steps 6 and 7 are used to model changes in dosimetry. (9) According to the results of step 8, the need for ART (replanning) is determined. If ART is needed then repeat from step 1; if not, treatment continues (step 5). Several authors have proposed algorithms for segmentation of breast tissues (4-7). A study that validates these methods would be useful. In the absence of true segmentation, validation becomes a challenging task. We have developed a 2-stage approach for evaluating tissue segmentation methods. The first stage evaluates segmentation algorithms using the knowledge that breast composition remains unchanged when imaged in different positions. This evaluates whole-breast segmentation but provides incomplete information and is susceptible to false-positive results (ie, 2 different breast segmentations resulting in a similar but wrong composition). The second stage validates tissue segmentation against expert outlines, which provides adequate validation but is time consuming and therefore feasible only for a limited number of images. Given the limitations of each, the 2-stage evaluation approach is proposed. We hypothesise that the accuracy of segmentation of tissues is affected by their distribution. Figure 2a and b illustrate 2
Fig. 1. Workflow of the proposed adaptive radiation therapy (ART) based on the modeling of tissue mechanics. CBCT Z cone beam computed tomography; FEM Z finite element methods; TB Z tumor bed; WB Z whole breast. fibroglandular tissue distributions (sparse and nonsparse). We tested whether the sparseness of the fibroglandular tissue distribution affects the accuracy of segmentation. The goals of this study were to investigate the accuracy of various breast tissue segmentation methods and determine the effects of tissue distribution on the segmentation accuracy for CT in breast cancer patients.
Methods and Materials This study validates segmentation methods from the literature and compares them using a dataset comprising matching prone and supine CT scans to select the most accurate segmentation method for breast modeling. The methods and their adjustable parameters to be optimized are listed in the Table 1. Four methods were used, from which 20 submethods were derived. All methods required some form of thresholding except k-means clustering. Methods were used to segment datasets into fibroglandular and fatty tissues, and in some cases also into background (a voxel that does not lie within the WB) and “other” (a mixture of fibroglandular and fatty tissues). The WB was segmented from the CT using clinician outlining, and then tissue segmentation methods were applied. A parameter called the fibroglandular composition (FC) was used to quantify the
Volume 84 Number 3 2012
Breast tissue segmentation for ART e421
Fig. 2. Sample middle-breast computed tomograpy images with whole-breast contour marked. (a) Breast with sparse distribution (expert rank Z 5) of fibroglandular tissue. (b) Breast with nonsparse distribution (expert rank Z 1) of fibroglandular tissue. fibroglandular composition of the breast. Fibroglandular composition is the percentage of breast that is fibroglandular tissue. The first step in evaluating these algorithms was the comparison of FC for the matching prone and supine data. It should be the same in both positions. The second step was comparison of segmentation with the tissue outlines from 3 experts.
Patients Patient data used in this study were originally collected to compare prone and supine positioning for breast radiation therapy (8, 9). Patients underwent lumpectomy, during which up to 6 pairs of titanium clips were placed in the TB. The patients underwent CT imaging in both positions (from cervical vertebra 6 to below the diaphragm) on the same day. The CT data comprised axial slices. On each slice WB clinical target volume (CTV) was delineated by a single clinician (A.K.). Example WB contours marked on single CT slices are shown in Fig. 2a and b. Kirby et al (8) described details of the definition of WB, patient positioning, and image acquisition. Datasets of 24 patients from that study were used in this work.
Interactive thresholding In this method the user interactively sets intensity CT number threshold. All the voxels with CT number higher than this threshold are labeled as fibroglandular and other voxels as fat, until the user judges the best segmentation is achieved.
K-means clustering This method results in clustering of data comprising N points into c classes (7). Segmentation was performed using 3 classes: background, fibroglandular, and fat.
Fuzzy c-means (FCM) clustering
Segmentation methods Physical density thresholding The calibration look-up table of the CT scanner was used to convert CT number to physical density. Physical density ranges
Table 1
for different breast tissues were obtained from the literature (10). Ranges for fatty and fibroglandular tissue were 0.917-0.939 g/cm3 and 1.013-1.047 g/cm3, respectively. Voxels with physical density within these ranges were classified as fat or fibroglandular tissue. We call this the hard threshold method. An expanded tissue range (the soft threshold method) was also investigated, in which voxels with physical density less than 0.939 g/cm3 were classified as fat, and all the voxels with physical density higher than 1.013 g/cm3 were classified as fibroglandular.
This method results in fuzzy clustering of the dataset (5). Fuzzy clustering means that the voxels are assigned probabilities of belonging to fat or fibroglandular tissues. After fuzzy clustering, some probability threshold has to be applied to yield the segmentation. Thresholding was carried out in 2 ways: using
Breast tissue segmentation methods and their associated parameters that were evaluated
Method number
Method name
Parameters*
1
Physical density thresholding
Physical density range
2 3
Interactive thresholding Fuzzy c-means clustering (FCM)
Not applicable Number of classes (or clusters) and the probability thresholdsy
4
k-means clustering
Not applicable
Submethod name
Submethod number
Hard threshold Soft threshold Interactive threshold FCM with 3 classes (FCM3) FCM with 4 classes (FCM4) k-means
1 2 3 4-11 12-19 20
* Denotes a quantity in each algorithm that is variable. y Different probability thresholds were the fibroglandular class at 0.10, 0.20, 0.40, 0.50, 0.60, 0.80, and 0.90; and the maximum probability threshold.
e422 Juneja et al.
International Journal of Radiation Oncology Biology Physics
various probabilities and maximum probability threshold. In the latter, a voxel was assigned to the class (fibroglandular, fat) for which it has highest membership value (probability of belonging). The effect of varying threshold was investigated. Segmentation was performed with 3 classes (FCM3) (ie, background, fibroglandular, and fat) and with 4 classes (FCM4) by adding another class labeled as “other.”
Zijdenbos et al (11) recommended that DSC >0.700 indicates a good spatial overlap between measurement pairs. jOXSj DSCZ2 ; ð3Þ jOj þ jSj
Evaluation of segmentation
Effect of tissue distribution on segmentation accuracy
The results of 20 segmentation submethods (Table 1) were generated in the investigation of the 4 methods. An evaluation scheme was followed to simplify the comparison of these methods. First, for all FCM the best FCM threshold values were found. Then, the best FCM and other segmentation methods were compared with each other to establish the most accurate method for tissue segmentation.
Prone and supine evaluation Segmentation methods were used to find the volume of fibroglandular tissues, and FC was calculated using Eq. 1 for all datasets. For each patient, relative difference in FC between prone and supine positions (DrFC ) was calculated using Eq. 2. FCZ100 VFG =Vtot
where S is the algorithmic segmentation, and O is the expert outline.
Breast datasets were visually assessed and classified by an expert (E.J.H.) for sparseness of fibroglandular tissue distribution. The expert ranked the distribution of fibroglandular tissue on a scale of 1-5 (where 5 is the most sparse, ie, very thin strands of fibroglandular tissue) using the Pinnacle (Philips Medical Systems, Bothwell, WA) radiation therapy planning system for visualization. The expert was blinded to the fibroglandular tissue segmentation results of the algorithms. Patient scans were divided into 2 groups according to the expert ranks: nonsparse (rank 1-3) and sparse (rank 4-5). Accuracy of the segmentation methods for these 2 groups was compared.
Statistical analysis
ð1Þ
abs FCpr FCsup ; Dr FCZ2) FCpr þ FCsup
ð2Þ
where VFG and Vtot are the volume of fibroglandular tissue and WB, respectively, and FCpr and FCsup are FC of prone and supine position, respectively.
A paired t test was used to compare the FC measured for the supine and prone data. Analysis of variance (ANOVA) followed by post hoc comparisons (Bonferroni, P<.05) were used to test for differences between methods. The Kruskal-Wallis test was used for nonparametric ANOVA, and the Friedman test was used for nonparametric repeated-measures ANOVA (12). The Wilcoxon rank Z sum test was used to test the significance of the differences between the 2 groups separated by tissue distribution rank.
Expert validation
Results Midbreast CT slices from supine scans of 12 of the patients were selected randomly and outlined by the 3 observers (1 radiologist and 2 radiation oncologists), yielding 36 manual outlines. Outlines were used individually for pairwise comparison with algorithmic segmentations to validate the algorithms. Experts were asked to mark fibroglandular tissue, seroma, and scar. Scar and seroma were combined with fibroglandular tissue for validation of methods because they have CT number similar to the fibroglandular tissue. The combined tissues were compared with fibroglandular tissue segmented by the above algorithms using the Dice similarity coefficient (DSC) given by Eq. 3. A value of DSC Z 1 indicates perfect agreement; a value of 0 indicates no overlap.
Table 2
Figure e1 (available online) gives an example of segmentation by the various methods and the experts.
Prone and supine evaluation The various segmentation methods gave different FC of the breasts (see mean FC [prone, supine] in Table 2 and Fig. 3a). Additionally, the breast composition varied among the patients, as seen by standard deviation around mean FC (prone, supine) in Table 2.
Mean fibroglandular composition (FC) and its relative difference from Eq. 2, and P value Method
Physical density, hard Physical density, soft Interactive thresholding FCM3 at threshold Z 0.20 k-means clustering
FC supine* 2% 5% 11% 32% 23%
Abbreviation: FCM3 Z fuzzy c-means with 3 classes. * Mean 1 standard deviation.
2% 6% 8% 5% 5%
FC prone* 2% 5% 11% 31% 23%
2% 5% 7% 5% 5%
Dr(FC)*
P
.51 .48 .51 .35 .41
15% 21% 13% 8% 10%
17% 21% 9% 7% 9%
Volume 84 Number 3 2012
Breast tissue segmentation for ART e423 between FC values measured from prone and supine scans were not significant (Table 2). The various methods gave different FC values, and the mean relative difference between prone and supine values from the respective methods was between 9% and 21%. The Kruskal-Wallis test did not indicate significant differences (PZ.08) between the methods. Fibroglandular composition measured by physical density thresholdings was smaller than with the other methods. Visual examination of segmented images against expert outlines confirmed that these methods were not identifying all the fibroglandular tissue, and therefore they were not considered for further investigation by expert validation.
Expert validation Dice similarity coefficient was calculated for FCM3, FCM4, interactive thresholding, and k-means clustering methods. Of the 36 slices outlined by experts, 16 (from 8 patients) had scar tissue as well as fibroglandular tissue marked, 20 had only fibroglandular tissue marked on them, and none had seroma. Generally there was good agreement between expert outlines; mean DSC for the pairs of expert outlines was 0.85 (standard deviation was 0.08).
Expert validation of FCM
Fig. 3. Comparison of fuzzy c-means with 3 classes (FCM3) and with 4 classes (FCM4) at various probability thresholds of the fibroglandular class. (a) Mean fibroglandular composition. (b) Mean Dice similarity coefficient. Error bars show 1 standard deviation.
Prone and supine evaluation of FCM Mean (averaged over all patients) FC measured using FCM3 and FCM4 decreases with an increase in threshold level (Fig. 3a). Fibroglandular composition decreased with increasing threshold because the fibroglandular class assignment was made more stringent. Fibroglandular composition was smaller for FCM4 than FCM3 at a given threshold because in FCM4 the membership values for fibroglandular class will always be less than or equal to that for FCM3. Mean FC (supine, prone) with maximum probability threshold was FCM3 (24%, 23%) and FCM4 (13%, 12%). Mean (averaged over all patients) relative difference between prone and supine FC for all probability thresholds was approximately 9% and 17% with 3 and 4 classes, respectively. The FCM3 was further compared with other segmentation methods.
Prone-supine evaluation and comparison of the various methods Fibroglandular composition values, corresponding standard deviations, and relative differences are given in Table 2. Differences
The validation of FCM segmentation with 3 classes (FCM3) and with 4 classes (FCM4) as a function of threshold is presented in Fig. 3b; it shows the mean (over all expert outlines) DSCs. Mean DSCs for FCM3 and FCM4 with the maximum probability threshold were 0.65 0.13 and 0.49 0.19, respectively. It was found that for any threshold, the average DSC for segmentation with FCM3 was higher than for FCM4. Friedman’s test (12) on the DSC of FCM3 and FCM4 methods with different probability thresholds indicated that FCM3 performed significantly better (P<.001) than FCM4. The FCM3 thresholded at 0.10, 0.15, and 0.20 had the highest mean DSC, which was 0.70. Friedman’s test on the DSC of FCM3 at different threshold values indicated that there were differences between different thresholding levels (P<.001), though post hoc analysis (Bonferroni, P<.05) did not give any pairwise significant differences.
Expert validation and comparison of the methods Dice similarity coefficient values for the various methods, namely FCM3 with thresholding of 0.10, 0.15, and 0.20, k-means, and interactive thresholding were compared. Mean DSCs of FCM3 with 3 thresholds, k-means, and interactive thresholding methods were 0.70, 0.67, and 0.61, respectively. The FCM3 methods with greatest mean DSC gave segmentation closest to expert delineation. Although Friedman’s test on DSC for these methods found no significant differences (P<.12) between them.
Effect of tissue distribution on segmentation accuracy Nine of the 24 patients in the study were classified as having a sparse tissue distribution according to the expert rank. Figure 4 summarizes the segmentation accuracy (DSC) measured for the 2 groups of patients dichotomized in terms of the sparseness of their fibroglandular tissues. All the segmentation algorithms had significantly (P<.005) higher DSCs for the nonsparse group than for the sparse group. Moreover, the agreement among experts outlines (experts overlap) for the nonsparse group was
e424 Juneja et al.
International Journal of Radiation Oncology Biology Physics
Fig. 4. Segmentation methods accuracy and experts overlap are plotted for 2 groups: nonsparse and sparse tissue distribution. There was significant (P<.005) difference between these 2 groups for all the methods and experts overlap. In each box, median is marked by central mark, the 25th and 75th percentile are the edges of the box, and the error bar represents the most extreme data points. Abbreviations are given in Table 1.
significantly (P<.005) more than for the sparse group. Mean relative difference in the FC between prone and supine (DrFC ) for nonsparse and sparse tissue groups were 6% and 12%, respectively, and the difference between the 2 groups was not significant (PZ.11).
Discussion In the present study, the results demonstrated that FCM3 thresholded at 0.10, 0.15, and 0.20 generated segmentation results closest to expert outlining. Similarly, Ertas et al (4) noted for breast MRI that the correlation between breast densities estimated on the basis of interactive thresholding and FCM3 is highest when a threshold of 0.20 is used. The evaluation of segmentation algorithms based on the assumption that the breast composition remains the same when measured in prone and supine CT scans did not show significant differences between the methods. In 7 of 24 cases, unexpectedly large differences (>10%) between prone and supine FC (DrFC ) were found. As our subsequent analysis found, the reason for the large relative difference in FC is probably the poor accuracy of breast tissue segmentation algorithms for sparse breasts. Otherwise our findings were compatible with those of Nie et al (6), who reported in their study of breast MRI a 3%-6% variation in measurements of FC with body positioning. We have shown for the first time that the sparseness of fibroglandular tissue significantly affects the accuracy of breast tissue segmentation. Adaptive radiation therapy based on modeling of tissue mechanics would not therefore currently be applicable to patients with this sparseness. We believe that this is an important result; to our knowledge the effect of the distribution of fibroglandular tissue on the accuracy of the segmentation algorithms has not previously been investigated. In this regard, a next step is to develop an automated analysis method to assess fibroglandular
tissue distribution and identify imaging data that are difficult to segment. Dose escalation to the region of the TB has been shown to improve local control in selected patients with early breast cancer (13). Local control benefits will be optimized (and effects on nontarget tissues minimized) by adapting the radiation therapy plan to changes in the TB and surrounding tissues. At present, whether boost doses are delivered sequentially or concomitantly, replanning of treatment based on a repeat CT is often required because of changes in the TB, sometimes incurring delays to treatment due to the need for reoutlining and replanning of radiation therapy. The aim of ART based on modeling of tissue mechanics is to deduce changes in TB through deformable registration of planning CT and CBCT (based on tissue mechanical properties) at each fraction and to predict the need for replanning. The registration of breast scans is nontrivial because breast comprises soft tissues and hence has little X-ray contrast, and the breast can undergo large deformation between different scans. Commonly used deformable registration methods do not account for physical characteristics of breast tissue (such as tissue types, tissue elasticity) and may lead to unrealistic deformation. Modeling of tissue mechanics using FEM has been shown to significantly improve the accuracy of deformable registration through the imposition of physically realistic constraints on breast deformation (14). Furthermore, the modeling of tissue mechanics could be used to describe the changes in TB with respect to breast and develop predictive models for ART and hence reduce the number of CBCTs required during treatment. In conclusion, this study has compared and validated breast tissue segmentation methods using prone and supine CTs and found that FCM3 gives the most accurate segmentation of breast tissue. The distribution of tissues within the breast affects the performance of segmentation methods. The understanding gained in this study is expected to help in using modeling of tissue mechanics for the purpose of adaptive breast radiation therapy.
Volume 84 Number 3 2012
References 1. Clarke M, Collins R, Darby S, et al. Effects of radiotherapy and of differences in the extent of surgery for early breast cancer on local recurrence and 15-year survival: an overview of the randomised trials. Lancet 2005;366:2087-2106. 2. Kim LH, Vicini F, Yan D. What do recent studies on lumpectomy cavity volume change imply for breast clinical target volumes? Int J Radiat Oncol Biol Phys 2008;72:1-3. 3. Harris EJ, Donovan EM, Yarnold JR, et al. Characterization of target volume changes during breast radiotherapy using implanted fiducial markers and portal imaging. Int J Radiat Oncol Biol Phys 2009;73: 958-966. 4. Ertas G, Reed S, Leach MO. Comparison of breast density assessments based on interactive thresholding and automated fast fuzzy c-means clustering in three-dimensional MR imaging. IFMBE Proc 2009;25:1893-1896. 5. Bezdek JC, Ehrlich R. FCM: the fuzzy c-means clustering algorithm. Comput Geosci 1984;10:191-203. 6. Nie K, Chen JH, Chan S, et al. Development of a quantitative method for analysis of breast density based on three-dimensional breast MRI. Med Phys 2008;35:5253-5262.
Breast tissue segmentation for ART e425 7. MacQueen JB. Some methods for classification and analysis of multivariate observations. Proc Fifth Berkeley Symp Math Stat Prob 1967;1:281-297. 8. Kirby AM, Evans PM, Donovan EM, et al. Prone versus supine positioning for whole and partial-breast radiotherapy: a comparison of non-target tissue dosimetry. Radiother Oncol 2010;96:178-184. 9. Kirby AM, Evans PM, Helyer SJ, et al. A randomised trial of Supine versus Prone breast radiotherapy (SuPr study): Comparing set-up errors and respiratory motion. Radiother Oncol 2011;100:221-226. 10. Johns PC, Yaffe MJ. X-ray characterisation of normal and neoplastic breast tissues. Phys Med Biol 1987;32:675-695. 11. Zijdenbos AP, Dawant BM, Margolin RA, et al. Morphometric analysis of white matter lesions transaction of medical imaging in MR images: method and validation. IEEE Trans Med Imaging 1994;13:716-724. 12. Hollander M, Wolfe DA. Nonparametric Statistical Methods. 2nd ed. Hoboken NJ: John Wiley & Sons; 1999. 13. Poortmans PM, Collette L, Horiot JC, et al. Impact of the boost dose of 10 Gy versus 26 Gy in patients with early stage breast cancer after a microscopically incomplete lumpectomy: 10-year results of the randomised EORTC boost trial. Radiother Oncol 2009;90:80-85. 14. Lee A, Schnabel J, Rajagopal V, et al. Breast image registration by combining finite elements and free-form deformations. Digital Mammogr 2010;736-743.
Radiation Oncology biology
physics
www.redjournal.org
Physics Contribution
Adaptive Breast Radiation Therapy Using Modeling of Tissue Mechanics: A Breast Tissue Segmentation Study Prabhjot Juneja, MSc,* Emma J. Harris, PhD,* Anna M. Kirby, FRCR,y and Philip M. Evans, DPhil* *Joint Department of Physics, Institute of Cancer Research, Sutton, United Kingdom; and yDepartment of Academic Radiotherapy, Royal Marsden National Health Service Foundation Trust, Sutton, United Kingdom Received Dec 8, 2011, and in revised form Apr 26, 2012. Accepted for publication May 6, 2012
Summary Modeling of tissue mechanics coupled with CBCT may be used to predict the requirement for replanning (ie, adaptive radiation therapy). To model patients’ breasts, component tissues need to be segmented and assigned material properties. This study validates tissue segmentation methods and evaluates the effect of tissue distribution on segmentation accuracy. We found the most accurate method to be fuzzy c-means clustering (3 classes) and that the segmentation accuracy is related to the distribution of fibroglandular tissue.
Purpose: To validate and compare the accuracy of breast tissue segmentation methods applied to computed tomography (CT) scans used for radiation therapy planning and to study the effect of tissue distribution on the segmentation accuracy for the purpose of developing models for use in adaptive breast radiation therapy. Methods and Materials: Twenty-four patients receiving postlumpectomy radiation therapy for breast cancer underwent CT imaging in prone and supine positions. The whole-breast clinical target volume was outlined. Clinical target volumes were segmented into fibroglandular and fatty tissue using the following algorithms: physical density thresholding; interactive thresholding; fuzzy c-means with 3 classes (FCM3) and 4 classes (FCM4); and k-means. The segmentation algorithms were evaluated in 2 stages: first, an approach based on the assumption that the breast composition should be the same in both prone and supine position; and second, comparison of segmentation with tissue outlines from 3 experts using the Dice similarity coefficient (DSC). Breast datasets were grouped into nonsparse and sparse fibroglandular tissue distributions according to expert assessment and used to assess the accuracy of the segmentation methods and the agreement between experts. Results: Prone and supine breast composition analysis showed differences between the methods. Validation against expert outlines found significant differences (P<.001) between FCM3 and FCM4. Fuzzy c-means with 3 classes generated segmentation results (mean DSC Z 0.70) closest to the experts’ outlines. There was good agreement (mean DSC Z 0.85) among experts for breast tissue outlining. Segmentation accuracy and expert agreement was significantly higher (P<.005) in the nonsparse group than in the sparse group. Conclusions: The FCM3 gave the most accurate segmentation of breast tissues on CT data and could therefore be used in adaptive radiation therapy-based on tissue modeling. Breast tissue segmentation methods should be used with caution in patients with sparse fibroglandular tissue distribution. Ó 2012 Elsevier Inc.
Reprint requests to: Prabhjot Juneja, MSc, Institute of Cancer Research, Joint Department of Physics, Sutton SM2 5PT, United Kingdom. Tel: (þ44) 208-661-3490; Fax: (þ44) 208-643-3812; E-mail: Prabhjot. [email protected] P.J. is supported by an Institute of Cancer Research PhD studentship. This work was supported by Cancer Research-UK under program grant reference no. C46/A2131. Int J Radiation Oncol Biol Phys, Vol. 84, No. 3, pp. e419ee425, 2012 0360-3016/$ - see front matter Ó 2012 Elsevier Inc. All rights reserved. doi:10.1016/j.ijrobp.2012.05.014
Conflict of interest: none. Supplementary material for this article can be found at www.redjournal.org. AcknowledgmentsdThe authors thank Galvin Polundniowski and Imogen Locke for valuable discussions, and Gigin Lin and Tarun Durga for tissue outlining.
e420 Juneja et al.
International Journal of Radiation Oncology Biology Physics
Introduction Adjuvant breast radiation therapy after breast-conserving surgery in early-stage breast cancer patients has been shown to reduce the risk of local recurrence (1). The region proximal to the excision cavity, the “tumor bed” (TB), is the most likely place for recurrence and often receives escalated dose. Many imaging studies have shown that the TB volume changes significantly over the time-frame relevant to radiation therapy planning and delivery (2, 3). Adaptive radiation therapy (ART) is used to correct for changes in the TB and involves repeat computed tomography (CT) and replanning of the treatment. The workload related to reimaging and replanning is high. Currently image guided breast radiation therapy uses portal imaging or cone beam CT (CBCT) to visualize bony anatomy or fiducial markers as surrogates for the TB. These techniques are inadequate for detecting the actual changes in TB and other breast tissues. In this work we use modeling of tissue changes based on the finite element method (FEM) to model tissue changes and information from the planning CT (higher quality than CBCT), to understand how the breast tissues deform and move. At each treatment fraction the planning CT is deformably registered to the CBCT using the FEM model to physically constrain the deformable registration. This allows more accurate quantification of TB changes and their effect on the treatment, thereby determining the requirement for replanning. This provides a means for adaptive breast radiation therapy. The proposed ART workflow is shown in Fig. 1 as 9 steps. (1) The patient has a planning CT. (2) An expert outlines whole breast (WB), TB, seroma, and scar on the images from step 1. (3) Fibroglandular and fatty tissues are segmented in the images using algorithms developed in this work. (4) Tissues are assigned mechanical properties, and a patient-specific FEM model is constructed from step 3. CT numbers of scar and seroma are similar to fibroglandular tissue and therefore are segmented together. The expert outlines of scar and seroma are used to separate them from fibroglandular tissues, and the deformation of these tissues can be deduced without the need for an additional planning CT scan. (5) The treatment starts, and at each treatment fraction the patient has CBCT. (6) The FEM model is used to register planning CT (step 1) and CBCT (step 5). (7) The deformation in step 6 is used to determine the TB deformation. (8) Tissue deformations in steps 6 and 7 are used to model changes in dosimetry. (9) According to the results of step 8, the need for ART (replanning) is determined. If ART is needed then repeat from step 1; if not, treatment continues (step 5). Several authors have proposed algorithms for segmentation of breast tissues (4-7). A study that validates these methods would be useful. In the absence of true segmentation, validation becomes a challenging task. We have developed a 2-stage approach for evaluating tissue segmentation methods. The first stage evaluates segmentation algorithms using the knowledge that breast composition remains unchanged when imaged in different positions. This evaluates whole-breast segmentation but provides incomplete information and is susceptible to false-positive results (ie, 2 different breast segmentations resulting in a similar but wrong composition). The second stage validates tissue segmentation against expert outlines, which provides adequate validation but is time consuming and therefore feasible only for a limited number of images. Given the limitations of each, the 2-stage evaluation approach is proposed. We hypothesise that the accuracy of segmentation of tissues is affected by their distribution. Figure 2a and b illustrate 2
Fig. 1. Workflow of the proposed adaptive radiation therapy (ART) based on the modeling of tissue mechanics. CBCT Z cone beam computed tomography; FEM Z finite element methods; TB Z tumor bed; WB Z whole breast. fibroglandular tissue distributions (sparse and nonsparse). We tested whether the sparseness of the fibroglandular tissue distribution affects the accuracy of segmentation. The goals of this study were to investigate the accuracy of various breast tissue segmentation methods and determine the effects of tissue distribution on the segmentation accuracy for CT in breast cancer patients.
Methods and Materials This study validates segmentation methods from the literature and compares them using a dataset comprising matching prone and supine CT scans to select the most accurate segmentation method for breast modeling. The methods and their adjustable parameters to be optimized are listed in the Table 1. Four methods were used, from which 20 submethods were derived. All methods required some form of thresholding except k-means clustering. Methods were used to segment datasets into fibroglandular and fatty tissues, and in some cases also into background (a voxel that does not lie within the WB) and “other” (a mixture of fibroglandular and fatty tissues). The WB was segmented from the CT using clinician outlining, and then tissue segmentation methods were applied. A parameter called the fibroglandular composition (FC) was used to quantify the
Volume 84 Number 3 2012
Breast tissue segmentation for ART e421
Fig. 2. Sample middle-breast computed tomograpy images with whole-breast contour marked. (a) Breast with sparse distribution (expert rank Z 5) of fibroglandular tissue. (b) Breast with nonsparse distribution (expert rank Z 1) of fibroglandular tissue. fibroglandular composition of the breast. Fibroglandular composition is the percentage of breast that is fibroglandular tissue. The first step in evaluating these algorithms was the comparison of FC for the matching prone and supine data. It should be the same in both positions. The second step was comparison of segmentation with the tissue outlines from 3 experts.
Patients Patient data used in this study were originally collected to compare prone and supine positioning for breast radiation therapy (8, 9). Patients underwent lumpectomy, during which up to 6 pairs of titanium clips were placed in the TB. The patients underwent CT imaging in both positions (from cervical vertebra 6 to below the diaphragm) on the same day. The CT data comprised axial slices. On each slice WB clinical target volume (CTV) was delineated by a single clinician (A.K.). Example WB contours marked on single CT slices are shown in Fig. 2a and b. Kirby et al (8) described details of the definition of WB, patient positioning, and image acquisition. Datasets of 24 patients from that study were used in this work.
Interactive thresholding In this method the user interactively sets intensity CT number threshold. All the voxels with CT number higher than this threshold are labeled as fibroglandular and other voxels as fat, until the user judges the best segmentation is achieved.
K-means clustering This method results in clustering of data comprising N points into c classes (7). Segmentation was performed using 3 classes: background, fibroglandular, and fat.
Fuzzy c-means (FCM) clustering
Segmentation methods Physical density thresholding The calibration look-up table of the CT scanner was used to convert CT number to physical density. Physical density ranges
Table 1
for different breast tissues were obtained from the literature (10). Ranges for fatty and fibroglandular tissue were 0.917-0.939 g/cm3 and 1.013-1.047 g/cm3, respectively. Voxels with physical density within these ranges were classified as fat or fibroglandular tissue. We call this the hard threshold method. An expanded tissue range (the soft threshold method) was also investigated, in which voxels with physical density less than 0.939 g/cm3 were classified as fat, and all the voxels with physical density higher than 1.013 g/cm3 were classified as fibroglandular.
This method results in fuzzy clustering of the dataset (5). Fuzzy clustering means that the voxels are assigned probabilities of belonging to fat or fibroglandular tissues. After fuzzy clustering, some probability threshold has to be applied to yield the segmentation. Thresholding was carried out in 2 ways: using
Breast tissue segmentation methods and their associated parameters that were evaluated
Method number
Method name
Parameters*
1
Physical density thresholding
Physical density range
2 3
Interactive thresholding Fuzzy c-means clustering (FCM)
Not applicable Number of classes (or clusters) and the probability thresholdsy
4
k-means clustering
Not applicable
Submethod name
Submethod number
Hard threshold Soft threshold Interactive threshold FCM with 3 classes (FCM3) FCM with 4 classes (FCM4) k-means
1 2 3 4-11 12-19 20
* Denotes a quantity in each algorithm that is variable. y Different probability thresholds were the fibroglandular class at 0.10, 0.20, 0.40, 0.50, 0.60, 0.80, and 0.90; and the maximum probability threshold.
e422 Juneja et al.
International Journal of Radiation Oncology Biology Physics
various probabilities and maximum probability threshold. In the latter, a voxel was assigned to the class (fibroglandular, fat) for which it has highest membership value (probability of belonging). The effect of varying threshold was investigated. Segmentation was performed with 3 classes (FCM3) (ie, background, fibroglandular, and fat) and with 4 classes (FCM4) by adding another class labeled as “other.”
Zijdenbos et al (11) recommended that DSC >0.700 indicates a good spatial overlap between measurement pairs. jOXSj DSCZ2 ; ð3Þ jOj þ jSj
Evaluation of segmentation
Effect of tissue distribution on segmentation accuracy
The results of 20 segmentation submethods (Table 1) were generated in the investigation of the 4 methods. An evaluation scheme was followed to simplify the comparison of these methods. First, for all FCM the best FCM threshold values were found. Then, the best FCM and other segmentation methods were compared with each other to establish the most accurate method for tissue segmentation.
Prone and supine evaluation Segmentation methods were used to find the volume of fibroglandular tissues, and FC was calculated using Eq. 1 for all datasets. For each patient, relative difference in FC between prone and supine positions (DrFC ) was calculated using Eq. 2. FCZ100 VFG =Vtot
where S is the algorithmic segmentation, and O is the expert outline.
Breast datasets were visually assessed and classified by an expert (E.J.H.) for sparseness of fibroglandular tissue distribution. The expert ranked the distribution of fibroglandular tissue on a scale of 1-5 (where 5 is the most sparse, ie, very thin strands of fibroglandular tissue) using the Pinnacle (Philips Medical Systems, Bothwell, WA) radiation therapy planning system for visualization. The expert was blinded to the fibroglandular tissue segmentation results of the algorithms. Patient scans were divided into 2 groups according to the expert ranks: nonsparse (rank 1-3) and sparse (rank 4-5). Accuracy of the segmentation methods for these 2 groups was compared.
Statistical analysis
ð1Þ
abs FCpr FCsup ; Dr FCZ2) FCpr þ FCsup
ð2Þ
where VFG and Vtot are the volume of fibroglandular tissue and WB, respectively, and FCpr and FCsup are FC of prone and supine position, respectively.
A paired t test was used to compare the FC measured for the supine and prone data. Analysis of variance (ANOVA) followed by post hoc comparisons (Bonferroni, P<.05) were used to test for differences between methods. The Kruskal-Wallis test was used for nonparametric ANOVA, and the Friedman test was used for nonparametric repeated-measures ANOVA (12). The Wilcoxon rank Z sum test was used to test the significance of the differences between the 2 groups separated by tissue distribution rank.
Expert validation
Results Midbreast CT slices from supine scans of 12 of the patients were selected randomly and outlined by the 3 observers (1 radiologist and 2 radiation oncologists), yielding 36 manual outlines. Outlines were used individually for pairwise comparison with algorithmic segmentations to validate the algorithms. Experts were asked to mark fibroglandular tissue, seroma, and scar. Scar and seroma were combined with fibroglandular tissue for validation of methods because they have CT number similar to the fibroglandular tissue. The combined tissues were compared with fibroglandular tissue segmented by the above algorithms using the Dice similarity coefficient (DSC) given by Eq. 3. A value of DSC Z 1 indicates perfect agreement; a value of 0 indicates no overlap.
Table 2
Figure e1 (available online) gives an example of segmentation by the various methods and the experts.
Prone and supine evaluation The various segmentation methods gave different FC of the breasts (see mean FC [prone, supine] in Table 2 and Fig. 3a). Additionally, the breast composition varied among the patients, as seen by standard deviation around mean FC (prone, supine) in Table 2.
Mean fibroglandular composition (FC) and its relative difference from Eq. 2, and P value Method
Physical density, hard Physical density, soft Interactive thresholding FCM3 at threshold Z 0.20 k-means clustering
FC supine* 2% 5% 11% 32% 23%
Abbreviation: FCM3 Z fuzzy c-means with 3 classes. * Mean 1 standard deviation.
2% 6% 8% 5% 5%
FC prone* 2% 5% 11% 31% 23%
2% 5% 7% 5% 5%
Dr(FC)*
P
.51 .48 .51 .35 .41
15% 21% 13% 8% 10%
17% 21% 9% 7% 9%
Volume 84 Number 3 2012
Breast tissue segmentation for ART e423 between FC values measured from prone and supine scans were not significant (Table 2). The various methods gave different FC values, and the mean relative difference between prone and supine values from the respective methods was between 9% and 21%. The Kruskal-Wallis test did not indicate significant differences (PZ.08) between the methods. Fibroglandular composition measured by physical density thresholdings was smaller than with the other methods. Visual examination of segmented images against expert outlines confirmed that these methods were not identifying all the fibroglandular tissue, and therefore they were not considered for further investigation by expert validation.
Expert validation Dice similarity coefficient was calculated for FCM3, FCM4, interactive thresholding, and k-means clustering methods. Of the 36 slices outlined by experts, 16 (from 8 patients) had scar tissue as well as fibroglandular tissue marked, 20 had only fibroglandular tissue marked on them, and none had seroma. Generally there was good agreement between expert outlines; mean DSC for the pairs of expert outlines was 0.85 (standard deviation was 0.08).
Expert validation of FCM
Fig. 3. Comparison of fuzzy c-means with 3 classes (FCM3) and with 4 classes (FCM4) at various probability thresholds of the fibroglandular class. (a) Mean fibroglandular composition. (b) Mean Dice similarity coefficient. Error bars show 1 standard deviation.
Prone and supine evaluation of FCM Mean (averaged over all patients) FC measured using FCM3 and FCM4 decreases with an increase in threshold level (Fig. 3a). Fibroglandular composition decreased with increasing threshold because the fibroglandular class assignment was made more stringent. Fibroglandular composition was smaller for FCM4 than FCM3 at a given threshold because in FCM4 the membership values for fibroglandular class will always be less than or equal to that for FCM3. Mean FC (supine, prone) with maximum probability threshold was FCM3 (24%, 23%) and FCM4 (13%, 12%). Mean (averaged over all patients) relative difference between prone and supine FC for all probability thresholds was approximately 9% and 17% with 3 and 4 classes, respectively. The FCM3 was further compared with other segmentation methods.
Prone-supine evaluation and comparison of the various methods Fibroglandular composition values, corresponding standard deviations, and relative differences are given in Table 2. Differences
The validation of FCM segmentation with 3 classes (FCM3) and with 4 classes (FCM4) as a function of threshold is presented in Fig. 3b; it shows the mean (over all expert outlines) DSCs. Mean DSCs for FCM3 and FCM4 with the maximum probability threshold were 0.65 0.13 and 0.49 0.19, respectively. It was found that for any threshold, the average DSC for segmentation with FCM3 was higher than for FCM4. Friedman’s test (12) on the DSC of FCM3 and FCM4 methods with different probability thresholds indicated that FCM3 performed significantly better (P<.001) than FCM4. The FCM3 thresholded at 0.10, 0.15, and 0.20 had the highest mean DSC, which was 0.70. Friedman’s test on the DSC of FCM3 at different threshold values indicated that there were differences between different thresholding levels (P<.001), though post hoc analysis (Bonferroni, P<.05) did not give any pairwise significant differences.
Expert validation and comparison of the methods Dice similarity coefficient values for the various methods, namely FCM3 with thresholding of 0.10, 0.15, and 0.20, k-means, and interactive thresholding were compared. Mean DSCs of FCM3 with 3 thresholds, k-means, and interactive thresholding methods were 0.70, 0.67, and 0.61, respectively. The FCM3 methods with greatest mean DSC gave segmentation closest to expert delineation. Although Friedman’s test on DSC for these methods found no significant differences (P<.12) between them.
Effect of tissue distribution on segmentation accuracy Nine of the 24 patients in the study were classified as having a sparse tissue distribution according to the expert rank. Figure 4 summarizes the segmentation accuracy (DSC) measured for the 2 groups of patients dichotomized in terms of the sparseness of their fibroglandular tissues. All the segmentation algorithms had significantly (P<.005) higher DSCs for the nonsparse group than for the sparse group. Moreover, the agreement among experts outlines (experts overlap) for the nonsparse group was
e424 Juneja et al.
International Journal of Radiation Oncology Biology Physics
Fig. 4. Segmentation methods accuracy and experts overlap are plotted for 2 groups: nonsparse and sparse tissue distribution. There was significant (P<.005) difference between these 2 groups for all the methods and experts overlap. In each box, median is marked by central mark, the 25th and 75th percentile are the edges of the box, and the error bar represents the most extreme data points. Abbreviations are given in Table 1.
significantly (P<.005) more than for the sparse group. Mean relative difference in the FC between prone and supine (DrFC ) for nonsparse and sparse tissue groups were 6% and 12%, respectively, and the difference between the 2 groups was not significant (PZ.11).
Discussion In the present study, the results demonstrated that FCM3 thresholded at 0.10, 0.15, and 0.20 generated segmentation results closest to expert outlining. Similarly, Ertas et al (4) noted for breast MRI that the correlation between breast densities estimated on the basis of interactive thresholding and FCM3 is highest when a threshold of 0.20 is used. The evaluation of segmentation algorithms based on the assumption that the breast composition remains the same when measured in prone and supine CT scans did not show significant differences between the methods. In 7 of 24 cases, unexpectedly large differences (>10%) between prone and supine FC (DrFC ) were found. As our subsequent analysis found, the reason for the large relative difference in FC is probably the poor accuracy of breast tissue segmentation algorithms for sparse breasts. Otherwise our findings were compatible with those of Nie et al (6), who reported in their study of breast MRI a 3%-6% variation in measurements of FC with body positioning. We have shown for the first time that the sparseness of fibroglandular tissue significantly affects the accuracy of breast tissue segmentation. Adaptive radiation therapy based on modeling of tissue mechanics would not therefore currently be applicable to patients with this sparseness. We believe that this is an important result; to our knowledge the effect of the distribution of fibroglandular tissue on the accuracy of the segmentation algorithms has not previously been investigated. In this regard, a next step is to develop an automated analysis method to assess fibroglandular
tissue distribution and identify imaging data that are difficult to segment. Dose escalation to the region of the TB has been shown to improve local control in selected patients with early breast cancer (13). Local control benefits will be optimized (and effects on nontarget tissues minimized) by adapting the radiation therapy plan to changes in the TB and surrounding tissues. At present, whether boost doses are delivered sequentially or concomitantly, replanning of treatment based on a repeat CT is often required because of changes in the TB, sometimes incurring delays to treatment due to the need for reoutlining and replanning of radiation therapy. The aim of ART based on modeling of tissue mechanics is to deduce changes in TB through deformable registration of planning CT and CBCT (based on tissue mechanical properties) at each fraction and to predict the need for replanning. The registration of breast scans is nontrivial because breast comprises soft tissues and hence has little X-ray contrast, and the breast can undergo large deformation between different scans. Commonly used deformable registration methods do not account for physical characteristics of breast tissue (such as tissue types, tissue elasticity) and may lead to unrealistic deformation. Modeling of tissue mechanics using FEM has been shown to significantly improve the accuracy of deformable registration through the imposition of physically realistic constraints on breast deformation (14). Furthermore, the modeling of tissue mechanics could be used to describe the changes in TB with respect to breast and develop predictive models for ART and hence reduce the number of CBCTs required during treatment. In conclusion, this study has compared and validated breast tissue segmentation methods using prone and supine CTs and found that FCM3 gives the most accurate segmentation of breast tissue. The distribution of tissues within the breast affects the performance of segmentation methods. The understanding gained in this study is expected to help in using modeling of tissue mechanics for the purpose of adaptive breast radiation therapy.
Volume 84 Number 3 2012
References 1. Clarke M, Collins R, Darby S, et al. Effects of radiotherapy and of differences in the extent of surgery for early breast cancer on local recurrence and 15-year survival: an overview of the randomised trials. Lancet 2005;366:2087-2106. 2. Kim LH, Vicini F, Yan D. What do recent studies on lumpectomy cavity volume change imply for breast clinical target volumes? Int J Radiat Oncol Biol Phys 2008;72:1-3. 3. Harris EJ, Donovan EM, Yarnold JR, et al. Characterization of target volume changes during breast radiotherapy using implanted fiducial markers and portal imaging. Int J Radiat Oncol Biol Phys 2009;73: 958-966. 4. Ertas G, Reed S, Leach MO. Comparison of breast density assessments based on interactive thresholding and automated fast fuzzy c-means clustering in three-dimensional MR imaging. IFMBE Proc 2009;25:1893-1896. 5. Bezdek JC, Ehrlich R. FCM: the fuzzy c-means clustering algorithm. Comput Geosci 1984;10:191-203. 6. Nie K, Chen JH, Chan S, et al. Development of a quantitative method for analysis of breast density based on three-dimensional breast MRI. Med Phys 2008;35:5253-5262.
Breast tissue segmentation for ART e425 7. MacQueen JB. Some methods for classification and analysis of multivariate observations. Proc Fifth Berkeley Symp Math Stat Prob 1967;1:281-297. 8. Kirby AM, Evans PM, Donovan EM, et al. Prone versus supine positioning for whole and partial-breast radiotherapy: a comparison of non-target tissue dosimetry. Radiother Oncol 2010;96:178-184. 9. Kirby AM, Evans PM, Helyer SJ, et al. A randomised trial of Supine versus Prone breast radiotherapy (SuPr study): Comparing set-up errors and respiratory motion. Radiother Oncol 2011;100:221-226. 10. Johns PC, Yaffe MJ. X-ray characterisation of normal and neoplastic breast tissues. Phys Med Biol 1987;32:675-695. 11. Zijdenbos AP, Dawant BM, Margolin RA, et al. Morphometric analysis of white matter lesions transaction of medical imaging in MR images: method and validation. IEEE Trans Med Imaging 1994;13:716-724. 12. Hollander M, Wolfe DA. Nonparametric Statistical Methods. 2nd ed. Hoboken NJ: John Wiley & Sons; 1999. 13. Poortmans PM, Collette L, Horiot JC, et al. Impact of the boost dose of 10 Gy versus 26 Gy in patients with early stage breast cancer after a microscopically incomplete lumpectomy: 10-year results of the randomised EORTC boost trial. Radiother Oncol 2009;90:80-85. 14. Lee A, Schnabel J, Rajagopal V, et al. Breast image registration by combining finite elements and free-form deformations. Digital Mammogr 2010;736-743.