Institute of Physics and Engineering in Medicine Phys. Med. Biol. 60 (2015) 4849–4871
Physics in Medicine & Biology doi:10.1088/0031-9155/60/12/4849
Measurement of prompt gamma profiles in inhomogeneous targets with a knife-edge slit camera during proton irradiation M Priegnitz1, S Helmbrecht1, G Janssens2, I Perali3,4, J Smeets2, F Vander Stappen2, E Sterpin5 and F Fiedler1 1
Helmholtz-Zentrum Dresden—Rossendorf, Institute of Radiation Physics, Bautzner Landstraße 400, 01328 Dresden, Germany 2 Ion Beam Applications SA, Chemin du Cyclotron 3, 1348 Louvain-la-Neuve, Belgium 3 Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria, Milano, Italy 4 Istituto Nazionale di Fisica Nucleare, Sezione di Milano, Milano, Italy 5 Center of Molecular Imaging, Radiotherapy and Oncology, Institut de recherche expérimentale et clinique, Université catholique de Louvain, Avenue Hippocrate 54, 1200 Brussels, Belgium E-mail: [email protected] Received 9 February 2015, revised 30 March 2015 Accepted for publication 24 April 2015 Published 9 June 2015 Abstract
Proton and ion beam therapies become increasingly relevant in radiation therapy. To fully exploit the potential of this irradiation technique and to achieve maximum target volume conformality, the verification of particle ranges is highly desirable. Many research activities focus on the measurement of the spatial distributions of prompt gamma rays emitted during irradiation. A passively collimating knife-edge slit camera is a promising option to perform such measurements. In former publications, the feasibility of accurate detection of proton range shifts in homogeneous targets could be shown with such a camera. We present slit camera measurements of prompt gamma depth profiles in inhomogeneous targets. From real treatment plans and their underlying CTs, representative beam paths are selected and assembled as one-dimensional inhomogeneous targets built from tissue equivalent materials. These phantoms have been irradiated with monoenergetic proton pencil beams. The accuracy of range deviation estimation as well as the detectability of range shifts is investigated in different scenarios. In most cases, range deviations can be detected within less than 2 mm. In close vicinity to low-density regions, range detection is challenging. In particular, a minimum beam penetration depth of 7 mm 0031-9155/15/124849+23$33.00 © 2015 Institute of Physics and Engineering in Medicine Printed in the UK
4849
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
beyond a cavity is required for reliable detection of a cavity filling with the present setup. Dedicated data post-processing methods may be capable of overcoming this limitation. Keywords: proton therapy, prompt gamma imaging, range verification, slit camera, sensitivity (Some figures may appear in colour only in the online journal) 1. Introduction Proton irradiation and ion irradiation gain more and more importance in tumour treatment as they yield advantageous physical dose distributions in comparison to conventional photon treatment. However, this dose distribution is very sensitive to changes in the irradiated volume. Minor deviations from the planned situation, for example, positioning inaccuracies or changes in tissue density due to swelling or filling of cavities, might result in considerable deviations from the desired dose distribution. Therefore, monitoring of the irradiation and especially of the particle range is desirable to fully exploit the advantages and the enhanced precision, as well as to avoid unwanted dose deviations. Presently, the only method in clinical application is particle therapy positron emission tomography (PT-PET) (Iseki et al 2003, Enghardt et al 2004, Parodi et al 2008, Hsi et al 2009, Nishio et al 2010, Zhu et al 2011, Bauer et al 2013, Knopf and Lomax 2013). This method is based on the detection of coincident annihilation photons resulting from the decay of β+active nuclei, which are produced during therapeutic irradiation. Current research activities on further range monitoring methods focus on the detection of secondary charged particles produced during ion beam treatment (interaction vertex imaging IVI (Henriquet et al 2012)) and on the detection of prompt gamma rays. These photons are also produced during therapeutic treatment; however, in contrast to PT-PET, they arise within femtoseconds after the impinging of the protons or ions. Different approaches for the detection of these prompt gamma rays are investigated worldwide, relying on the measurement of the spatial (Prompt Gamma Imaging PGI), energy (Prompt Gamma Spectroscopy PGS (Verburg and Seco 2014)), or the temporal distribution (Prompt Gamma Timing PGT (Golnik et al 2014)) of the prompt gammas. For the spatial measurement of the gamma distribution, a Compton Camera can be used, which is based on electronic collimation of the photons reaching the detector (Frandes et al 2010, Kormoll et al 2011, Richard et al 2011, Robertson et al 2011, Park et al 2012, Trovato et al 2013, Hueso-González et al 2014, Krimmer et al 2015, Thirolf et al 2014). Another possibility is the passive collimation of the gamma rays by means of a mechanical collimator made from, e.g. tungsten or lead. Research on different designs is performed, including cameras with pinhole collimators (Kim et al 2009), knife-edge slit collimators (Bom et al 2012, Smeets et al 2012), or multi-parallel slit collimators (Min et al 2012, Pinto et al 2014) for restricting the photon direction. In our work, we focus on the measurement of prompt gamma distributions with the knife-edge-shaped slit camera described by Smeets et al (2012) and Perali et al (2014). The performance of this camera was demonstrated in homogeneous targets by means of simulations and experiments. In the present work, the effects of inhomogeneous targets and the detection efficiency of range changes caused by either anatomical changes or small changes in the particle energy are investigated systematically. 4850
HU
0
Muscle
Lung
0
Soft tissue
Muscle
Adipose Air Mask
Air
Brain
Mask Air Adipose
HU
500
Brain
500 1000
Lung Tumour
Adipose
1500
1000
Bone
Bone
2000
Adipose
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
-500
-500 -1000 -1000 Protons -1500
Protons -1500 0
50
100
150 x / mm
200
250
0
100
200
300 x / mm
400
Figure 1. Profile through the head (left) and lung region (right). Hounsfield units (HU) of the real patient CT are given in black. The blue line shows the idealized profile. The beam is intended to come from the left, as in the schematic drawing of the delivered dose (orange line).
Figure 2. 1D heterogeneous target with tissue-equivalent inserts. Left: head target.
The hollow cylinder is not closed for better visibility. Right: lung target, completely closed, it is used in the experiment. In both pictures the beam comes from the left.
2. Materials and methods 2.1. Target definition and preparation
For our investigations with inhomogeneous targets, we defined different scenarios of interest for range verification in proton therapy. We chose an irradiation close to the skullbase with an air-filled cavity in the beam path and lung irradiation. Using real patient CT data and corresponding treatment plans, we selected representative paths of the beam through the tissue and made an idealization of the respective Hounsfield units (HU) (figure 1). In the lung case, we manually added a tumour region within the lung region. Tissues were assigned according to the HU under anatomical considerations. For practical implementation, we designed a 1D heterogeneous target. We used a cylindrical polymethyl methacrylate (PMMA) target consisting of two hollow half cylinders that can be filled with disks of different thicknesses and various tissue-equivalent materials (from Gammex-RMI GmbH, Germany) such as lung-, muscle-, or fat-equivalent materials (figure 2). 4851
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Table 1. Density ρ and chemical composition (mass fraction w in %) of the tissue equivalent Gammex materials (Gammex RMI GmbH, Germany 2015).
Material
ρ/g cm−3
w (H)
w (C)
w (N)
w (O)
w (Mg)
w (Si) w (Cl)
Lung Adipose Muscle Bone
0.30 0.92 1.05 1.82
8.46 9.06 8.10 3.41
59.38 72.30 67.17 31.41
1.96 2.25 2.42 1.84
18.14 16.27 19.85 36.50
11.19
0.78
0.10 0.13 0.14 0.04
w (Ca)
2.32 26.81
The target has a length of 40 cm, an outer diameter of 15 cm, and an inner diameter of 5 cm for insertions of the disks. It has to be mentioned that the Gammex materials are tissue-equivalent in terms of photon attenuation and, additionally, densities and proton ranges mimic human tissues well. However, chemical composition of these materials and the relative amount of elements differ with respect to biological tissue. Mimicking head irradiation, we assembled 10 mm of PMMA, representing the mask for patient positioning, followed by 5 mm air, 10 mm adipose tissue, i.e. the skin, 10 mm bone, 38 mm brain represented with PMMA, 10 mm air cavity, and an additional 38 mm brain represented by PMMA (due to a lack of brain-equivalent material). To ensure high accuracy in the size of the air cavities, they were realized as PMMA rings with a wall thickness of 3 mm and a well-defined thickness in beam direction. During irradiation of this target, the beam is intended to stop at a certain distance beyond the cavity. This would require an additional range shifter in front of the target, because the requested energy is rather low and cannot be provided directly by the proton therapy system. Therefore, we added 40 mm PMMA at the front face of the target. Irradiations in the head bear the risk of particle under-range due to cavity filling, which might happen due to swelling or mucus-filling concomitant to a cold. This effect was investigated by replacing the 10 mm air cavity with PMMA. For lung irradiation, we assembled 20 mm adipose, 50 mm muscle, 30 mm lung, and another 20 mm muscle tissue representing the tumour, followed by 30 mm lung, and, finally, we filled with muscle. The beam is intended to stop in the tumour, i.e. the part with 20 mm muscle-equivalent tissue. A problem in lung irradiation aside from the potential motion of the tumour is the need for the beam to cross the rib cage. Due to slight patient mispositioning or respiratory motion, a pencil beam might go through a rib, although it was planned to pass in between them through soft tissue or vice versa. To mimic such an accidentally irradiated rib, we replaced the 50 mm muscle part in the beginning of the target, in accordance with the present patient CT, with 40 mm muscle followed by 10 mm bone tissue. A photograph of both targets, head and lung, is shown in figure 2. The composition of the tissue-equivalent Gammex materials is given in table 1. 2.2. Irradiation
Experiments were performed at the Proton Therapy Center Czech in Prague, Czech Republic. The target was positioned on the patient table with the inhomogeneities along the beam axis and the expected Bragg peak position at the isocenter. Irradiation was performed with monoenergetic proton pencil beams along the target axis and without scanning. Different range modification scenarios were investigated. In the case of the head target, the beam energy was chosen so that the beam stops 5 mm (137.5 MeV) and 15 mm (144 MeV) beyond the cavity. In both cases, we imitated the effect of range shifts (which might have different reasons such as mispositioning or density changes) by changing the beam energy in 10 steps of 1 MeV 4852
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 3. Left: detector of the slit camera. Right: full slit camera with trolley and knife-
edge-shaped slit collimator.
each to higher and lower ranges, resulting in 21 different irradiations for each case. The same was done with the filled cavity target. In case of the lung irradiation, we chose the beam to stop in the middle of the tumour, i.e. beam energy was 110 MeV. To study the effect of accidentally hitting a rib, we irradiated the rib target with the same energy. For both targets, lung with and without rib, we also varied the beam energy from 100 MeV to 120 MeV in 1 MeV steps for investigation of range deviation effects. In each irradiation scenario, approximately 2.6 · 1011 protons were delivered. 2.3. Camera system
The detection of prompt gamma radiation was performed with a knife-edge-shaped slit camera, as described in detail in Perali et al (2014). The detector consists of two rows of 20 LYSO slabs, which are read-out by silicon photomultipliers (SiPM). The detector is mounted onto a trolley, which also holds the tungsten knife-edge slit collimator, figure 3. The distance between beam axis and collimator was set to 25 cm and the distance between detector and collimator was set to 20 cm, resulting in a magnification factor of 0.8. The two rows of 20 LYSO slabs with a crystal width of 4 mm each thereby produced a field of view (FOV) of 10 cm along the beam axis. The center of this FOV corresponding to the slit aperture of the collimator was aligned at the expected range of the impinging protons. The distance between the collimator front face and the target surface is 15.5 cm. A schematic overview on the experimental setup is shown in figure 4. Simulations, measurements, and calculations related to these experiments indicate that the camera exhibits a point spread function between 15 mm and 30 mm full width at half maximum (FWHM), depending on the geometrical setup and the size of the phantom. In the configuration used here, FWHM of the point spread function is approximately 20 mm. Before starting the experiments, energy calibration was performed as proposed by Perali (2014): first with 137Cs (662 keV) and 60Co (1173 keV and 1332 keV), then with the 1.37 MeV and 2.75 MeV gamma lines going along with the decay of 24Na (which was obtained from irradiation of an aluminum sample with protons according to 27Al(p,n + 3p)24Na), and finally 4853
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 4. Schematic overview of the experimental setup. The drawing is not true to
scale. A few prompt gamma photons are shown as green lines.
with the double escape peaks from excited 12C and 16O (3.42 MeV and 5.11 MeV) during proton irradiation of water. These energy calibration values have been fitted with an exponential characteristic to take into account the saturation of the SiPMs microcells for the high amount of light that arises at high energies. On the basis of this calibration, the lower and upper energy thresholds for the measurement of prompt gamma detection profiles were set to 3 MeV and 6 MeV, respectively. Only prompt gammas detected in between these thresholds are scored, because simulations by Smeets et al (2012) and measurements by Perali et al (2014) show that the prompt gamma rays in this energy window exhibit the best correlation with the beam penetration depth for the present camera design. Lower and higher energy gammas suffer from more contamination by uncorrelated events and are therefore discarded. 2.4. Simulation
The performed experiments were simulated with an analytical implementation based on a library of profiles pre-computed by the Monte Carlo PENH algorithm (Sterpin et al 2013). By means of a continuous slowing-down approximation (CSDA), the range of the protons and, consequently, the expected range deviation are obtained. Furthermore, to allow for a better interpretation of the data, the resulting prompt gamma emission as well as the expected detected profile have been simulated. As shown previously in measurements and full Monte Carlo simulation (Smeets et al 2012), the uncorrelated background signal detected by the camera as a result of neutron emissions and their secondaries appears to have an overall rather flat contribution to the detection profile. Thus, it is not simulated in this investigation. The materials used for the experiment, i.e. Gammex tissue-equivalent materials, were modelled very carefully in terms of their elemental composition. However, uncertainties in the ionization potential of those materials remain and affect simulations in terms of quantitative aspects. Therefore, the resulting particle ranges have an accuracy not better than 1 mm. A direct, quantitative comparison between measurement and simulation is beyond the scope of this article and is presented elsewhere (Sterpin et al 2015). 4854
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 5. Procedure of distal edge detection for two different measured profiles (left
and middle) with the help of their first and second derivatives. The dash–dotted vertical lines indicate the detected distal edge determined as described in (ii) in section 2.5. Right: for range deviation estimation, the two parts detected as in the left and middle parts of the figure are normalized and shifted against each other to minimize the area between them. This area is enclosed from the two curves and the dashed lines.
2.5. Data analysis
With the slit camera described one obtains one-dimensional prompt gamma profiles. The measurements consist of 20 single values, one for each 4 mm wide detector bin (i.e. the sum of two crystal slabs: the one from the lower and the one from the upper row). For analysis, a linear interpolation between these values is performed so that one point per millimeter in the target is obtained. A Gaussian function with 20 mm FWHM is applied to smooth the data and reduce statistical noise. Because the width of the smoothing function is not larger than the camera’s point spread function, it is expected that only patterns with spatial frequencies due to statistical fluctuations are reduced, providing no information regarding the real emission profiles. Evaluation of the range shift relies on the investigation of the distal edge of the measured profile. For the prompt gamma measurements presented here, different methods of distal edge detection were investigated. The best procedure here is found to be the following (figure 5): (i) Normalization and background subtraction: all profiles are scaled so that the mean of the three most proximal data points is one and the mean of the three most distal points is zero. (ii) Distal edge definition: it is defined as the most distal part of the profile, where data points are larger than a certain threshold (here, we chose 0.1 to suppress statistical background fluctuations), the profile is decreasing (first derivative 10 mm) in this target the accuracy of range shift estimation suffers slightly (errors up to 2 mm). This can be explained by the applied normalization to the profiles described here and originates from the increasingly different shape of the profiles compared to the reference. In case of the head target with empty cavity, the achievable accuracy depends very much on the distance between the cavity and the planned beam penetration depth. Two different scenarios have been studied. In one case, the beam is intended to stop 5 mm beyond the cavity (137.5 MeV proton energy), and in the other one the beam is to stop 15 mm beyond the cavity (144 MeV proton energy). Figure 8 shows simulations of emitted and detected gammas together with the HU distribution of the target for the empty cavity case. It can be seen that 4859
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
at 137.5 MeV, although the pencil beam stepped across the cavity, there is no corresponding information in the detection profile. Prompt gamma emissions reach their maximum at 25 MeV proton energy, which corresponds to approximately 6 mm residual proton range in water, and then starts declining very steeply to zero at approximately 6 MeV (i.e. 0.5 mm in water). Thus, there is basically no prompt gamma signal in the last 3 mm of the proton path. When the beam stops 5 mm beyond a 10 mm cavity, only approximately 2 mm of tissue beyond the cavity deliver a prompt gamma signal, whereas virtually no signal comes from the cavity itself. This would be measurable with a perfect camera. However, with our real camera the emission profile can be seen as multiplied by a factor of 4 · 10−4 for the efficiency, a point spread function of 20 mm FWHM, and a 1 : 1 signal-to-noise ratio. Therefore, the profile of 137.5 MeV contains almost no information on the cavity that was crossed by the pencil beam. In contrast, the 144 MeV profile, i.e. when the beam stops 15 mm beyond the cavity, shows evidence that the pencil beam crossed the cavity. In this case, overshoots can be detected with excellent accuracy as in the filled cavity cases. Undershoots, however, are recognized by the camera with even shorter range than their actual value, which is related to the missing signal from the cavity. Those points are detected with the actual range shift minus the cavity size and are therefore on a straight line parallel to and below the one representing the equality between estimated and expected shifts (figure 7). In the case of the beam stopping 5 mm beyond the cavity (137.5 MeV), undershoots are underestimated and overshoots are overestimated. Most data points in figure 7 (left) are on a straight line that follows the equation ‘detected range shift = actual range shift + cavity thickness’, i.e. parallel above the line representing equality between estimated and actual shifts. That means the size of the cavity is inherently included for the detected range shift because the reference profile contains almost no information that the cavity was actually stepped across by the pencil beam. Only in a very small region around the planned beam penetration depth (±2 MeV) can a sufficiently high accuracy be realized, because here the detection profile is also missing evidence that the cavity was stepped across. In the investigated lung target, a rather high accuracy for range deviation detection is found as long as the beam stop is in the tumour. This is indicated with black symbols in the respective part of figure 7. As soon as the beam range does not reach the tumour volume anymore, accuracy declines enormously (blue symbols in the right part of figure 7). The correlation between prompt gamma production and proton range seems to be lost in the lung tissue due to its low density compared to the surrounding tissue. 3.1.2. Detectability. As discussed, differences between detected prompt gamma shift and actual existing shift in particle range can occur depending on the composition of the target. The question to be answered now is, to what extent can the shift be detected even though the accuracy of the range deviation estimation can be weak in the presence of large density gradients? To answer this question, ROC curves are used. In figure 9 the histograms of detected shifts for the empty cavity head irradiation with 144 MeV and a shift of −5 MeV and +5 MeV, respectively, are shown for different numbers of incident protons. This energy shift results in a range shift of −8 mm and +8 mm, respectively. For the full statistic profiles (2.6 · 1011 protons) measured in the experiment, a good accuracy of less than 2 mm deviation from the real particle range shift was achieved for irradiation with larger energy (overshoot). However, for irradiation with lower energy (undershoot) the accuracy of shift estimation was rather poor, with a deviation of up to 10 mm from the applied shift due to the increasing influence of the cavity, as can be seen from figure 7 and as was discussed in the previous section. Nevertheless, in figure 9 it can be seen that in both cases, positive as well as negative shifts, the histograms can be separated quite well. This leads to comparatively high and very similar sensitivity and specificity 4860
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
(100 %, 100 %)
(100 %, 100 %) (98.4 %, 99.5 %)
(95.9 %, 96.3 %) (76.6 %, 67.0%)
(74.8 %, 75.0 %)
Figure 9. Top: histograms of detected shifts for the head target (empty cavity) with 144
MeV protons (green) and an energy shift (red) of −5 MeV (left, results in −8 mm range shift) and +5 MeV (right, results in +8 mm range shift), respectively. The distributions are shown for different numbers of incident protons. Bottom: ROC curves for these histograms. The inserted points show the points of maximized Youden index with sensitivity and specificity given in brackets. Note that the ROC curve for 109 protons is hidden below the one for 1010 protons.
for −5 MeV and +5 MeV shift, although the accuracy in shift estimation is very different. For irradiation with 107 protons, a sensitivity of approximately 75% and 77% and a specificity of approximately 75% and 67% (−5 MeV and +5 MeV, resp.) can be achieved for that special anatomical case. For 108 protons, more than 95% sensitivity and specificity can already be reached. That means that the detectability of the shift is equal for both cases, i.e. the shift can be detected with the same probability; however, the accuracy of the estimated shift is very different. The results of all energy shifts in terms of sensitivity and specificity are summarized in figure 10. As expected, for minor shifts in energy and, thus, in range, sensitivity and specificity are comparatively low, because the histograms of detected shifts are difficult to separate for the shifted and non-shifted case. For an increasing number of incident protons, an improvement can be seen. For the shift detection in the lung irradiation exemplary histograms for −2 MeV and +2 MeV shift are shown in figure 11. This shift in energy results in a range shift of −3 mm and +3 mm, respectively. It can be seen that for a lower number of protons, both histograms become difficult to distinguish. The results on the accuracy presented show that good accuracy of less than 2 mm can be achieved when statistics are very high. However, the results of the 4861
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 10. Sensitivity and specificity for the detection of range shifts in the head target
(empty cavity) due to variations in the proton energy around the nominal energy of 144 MeV.
analysis of ROC curves show that sensitivity and specificity decrease severely when lowering prompt gamma statistics, i.e. decreasing the number of incident protons. Results concerning sensitivity and specificity for all investigated shifts in energy are shown in figure 12. For irradiation with 108 protons or more, rather good sensitivity and specificity of more than 70% can be achieved in most cases. Although the accuracy in the range shift estimation is not satisfying for large shifts due to the missing correlation between proton range and gamma emission in low-density tissue (figure 7), the ROC curves show that the existence of the shift can be generally detected. Thus, further examinations on the source of the range deviations related with their potential consequences on the treatment delivery can be initiated. 3.2. Range deviation due to cavity filling or a rib 3.2.1. Accuracy. Another question to address is the accuracy and detectability of range devia-
tions due to cavity filling in the head target or accidental irradiation of a rib in lung irradiation. Therefore, the measured profiles of the head target with empty cavity and the lung target without rib have been taken as references. For range deviation analysis, the profile from the target with filled cavity and the target with rib, respectively, with the same proton energy have been used. The results are summarized in figure 13. Here, the detected range deviations are plotted as symbols for each proton energy. The expected range shift from simulation is given as a grey solid line. In the case of head irradiation (figure 13 left), data from two different setups are available. In the 137.5 MeV setup, the slit of the collimator was positioned at the range of the 137.5 MeV protons, and then the proton energy was changed ± 10 MeV in 1 MeV steps. The other setup is the 144 MeV setup with the slit of the collimator at the range of the 144 MeV protons. Thus, there is an overlap in proton energy from both measurement series (134 MeV–147 MeV). However, detected shifts do not coincide completely in that overlap due to a different target region within the FOV of the camera in both setups. For the low energies (
Physics in Medicine & Biology doi:10.1088/0031-9155/60/12/4849
Measurement of prompt gamma profiles in inhomogeneous targets with a knife-edge slit camera during proton irradiation M Priegnitz1, S Helmbrecht1, G Janssens2, I Perali3,4, J Smeets2, F Vander Stappen2, E Sterpin5 and F Fiedler1 1
Helmholtz-Zentrum Dresden—Rossendorf, Institute of Radiation Physics, Bautzner Landstraße 400, 01328 Dresden, Germany 2 Ion Beam Applications SA, Chemin du Cyclotron 3, 1348 Louvain-la-Neuve, Belgium 3 Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria, Milano, Italy 4 Istituto Nazionale di Fisica Nucleare, Sezione di Milano, Milano, Italy 5 Center of Molecular Imaging, Radiotherapy and Oncology, Institut de recherche expérimentale et clinique, Université catholique de Louvain, Avenue Hippocrate 54, 1200 Brussels, Belgium E-mail: [email protected] Received 9 February 2015, revised 30 March 2015 Accepted for publication 24 April 2015 Published 9 June 2015 Abstract
Proton and ion beam therapies become increasingly relevant in radiation therapy. To fully exploit the potential of this irradiation technique and to achieve maximum target volume conformality, the verification of particle ranges is highly desirable. Many research activities focus on the measurement of the spatial distributions of prompt gamma rays emitted during irradiation. A passively collimating knife-edge slit camera is a promising option to perform such measurements. In former publications, the feasibility of accurate detection of proton range shifts in homogeneous targets could be shown with such a camera. We present slit camera measurements of prompt gamma depth profiles in inhomogeneous targets. From real treatment plans and their underlying CTs, representative beam paths are selected and assembled as one-dimensional inhomogeneous targets built from tissue equivalent materials. These phantoms have been irradiated with monoenergetic proton pencil beams. The accuracy of range deviation estimation as well as the detectability of range shifts is investigated in different scenarios. In most cases, range deviations can be detected within less than 2 mm. In close vicinity to low-density regions, range detection is challenging. In particular, a minimum beam penetration depth of 7 mm 0031-9155/15/124849+23$33.00 © 2015 Institute of Physics and Engineering in Medicine Printed in the UK
4849
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
beyond a cavity is required for reliable detection of a cavity filling with the present setup. Dedicated data post-processing methods may be capable of overcoming this limitation. Keywords: proton therapy, prompt gamma imaging, range verification, slit camera, sensitivity (Some figures may appear in colour only in the online journal) 1. Introduction Proton irradiation and ion irradiation gain more and more importance in tumour treatment as they yield advantageous physical dose distributions in comparison to conventional photon treatment. However, this dose distribution is very sensitive to changes in the irradiated volume. Minor deviations from the planned situation, for example, positioning inaccuracies or changes in tissue density due to swelling or filling of cavities, might result in considerable deviations from the desired dose distribution. Therefore, monitoring of the irradiation and especially of the particle range is desirable to fully exploit the advantages and the enhanced precision, as well as to avoid unwanted dose deviations. Presently, the only method in clinical application is particle therapy positron emission tomography (PT-PET) (Iseki et al 2003, Enghardt et al 2004, Parodi et al 2008, Hsi et al 2009, Nishio et al 2010, Zhu et al 2011, Bauer et al 2013, Knopf and Lomax 2013). This method is based on the detection of coincident annihilation photons resulting from the decay of β+active nuclei, which are produced during therapeutic irradiation. Current research activities on further range monitoring methods focus on the detection of secondary charged particles produced during ion beam treatment (interaction vertex imaging IVI (Henriquet et al 2012)) and on the detection of prompt gamma rays. These photons are also produced during therapeutic treatment; however, in contrast to PT-PET, they arise within femtoseconds after the impinging of the protons or ions. Different approaches for the detection of these prompt gamma rays are investigated worldwide, relying on the measurement of the spatial (Prompt Gamma Imaging PGI), energy (Prompt Gamma Spectroscopy PGS (Verburg and Seco 2014)), or the temporal distribution (Prompt Gamma Timing PGT (Golnik et al 2014)) of the prompt gammas. For the spatial measurement of the gamma distribution, a Compton Camera can be used, which is based on electronic collimation of the photons reaching the detector (Frandes et al 2010, Kormoll et al 2011, Richard et al 2011, Robertson et al 2011, Park et al 2012, Trovato et al 2013, Hueso-González et al 2014, Krimmer et al 2015, Thirolf et al 2014). Another possibility is the passive collimation of the gamma rays by means of a mechanical collimator made from, e.g. tungsten or lead. Research on different designs is performed, including cameras with pinhole collimators (Kim et al 2009), knife-edge slit collimators (Bom et al 2012, Smeets et al 2012), or multi-parallel slit collimators (Min et al 2012, Pinto et al 2014) for restricting the photon direction. In our work, we focus on the measurement of prompt gamma distributions with the knife-edge-shaped slit camera described by Smeets et al (2012) and Perali et al (2014). The performance of this camera was demonstrated in homogeneous targets by means of simulations and experiments. In the present work, the effects of inhomogeneous targets and the detection efficiency of range changes caused by either anatomical changes or small changes in the particle energy are investigated systematically. 4850
HU
0
Muscle
Lung
0
Soft tissue
Muscle
Adipose Air Mask
Air
Brain
Mask Air Adipose
HU
500
Brain
500 1000
Lung Tumour
Adipose
1500
1000
Bone
Bone
2000
Adipose
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
-500
-500 -1000 -1000 Protons -1500
Protons -1500 0
50
100
150 x / mm
200
250
0
100
200
300 x / mm
400
Figure 1. Profile through the head (left) and lung region (right). Hounsfield units (HU) of the real patient CT are given in black. The blue line shows the idealized profile. The beam is intended to come from the left, as in the schematic drawing of the delivered dose (orange line).
Figure 2. 1D heterogeneous target with tissue-equivalent inserts. Left: head target.
The hollow cylinder is not closed for better visibility. Right: lung target, completely closed, it is used in the experiment. In both pictures the beam comes from the left.
2. Materials and methods 2.1. Target definition and preparation
For our investigations with inhomogeneous targets, we defined different scenarios of interest for range verification in proton therapy. We chose an irradiation close to the skullbase with an air-filled cavity in the beam path and lung irradiation. Using real patient CT data and corresponding treatment plans, we selected representative paths of the beam through the tissue and made an idealization of the respective Hounsfield units (HU) (figure 1). In the lung case, we manually added a tumour region within the lung region. Tissues were assigned according to the HU under anatomical considerations. For practical implementation, we designed a 1D heterogeneous target. We used a cylindrical polymethyl methacrylate (PMMA) target consisting of two hollow half cylinders that can be filled with disks of different thicknesses and various tissue-equivalent materials (from Gammex-RMI GmbH, Germany) such as lung-, muscle-, or fat-equivalent materials (figure 2). 4851
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Table 1. Density ρ and chemical composition (mass fraction w in %) of the tissue equivalent Gammex materials (Gammex RMI GmbH, Germany 2015).
Material
ρ/g cm−3
w (H)
w (C)
w (N)
w (O)
w (Mg)
w (Si) w (Cl)
Lung Adipose Muscle Bone
0.30 0.92 1.05 1.82
8.46 9.06 8.10 3.41
59.38 72.30 67.17 31.41
1.96 2.25 2.42 1.84
18.14 16.27 19.85 36.50
11.19
0.78
0.10 0.13 0.14 0.04
w (Ca)
2.32 26.81
The target has a length of 40 cm, an outer diameter of 15 cm, and an inner diameter of 5 cm for insertions of the disks. It has to be mentioned that the Gammex materials are tissue-equivalent in terms of photon attenuation and, additionally, densities and proton ranges mimic human tissues well. However, chemical composition of these materials and the relative amount of elements differ with respect to biological tissue. Mimicking head irradiation, we assembled 10 mm of PMMA, representing the mask for patient positioning, followed by 5 mm air, 10 mm adipose tissue, i.e. the skin, 10 mm bone, 38 mm brain represented with PMMA, 10 mm air cavity, and an additional 38 mm brain represented by PMMA (due to a lack of brain-equivalent material). To ensure high accuracy in the size of the air cavities, they were realized as PMMA rings with a wall thickness of 3 mm and a well-defined thickness in beam direction. During irradiation of this target, the beam is intended to stop at a certain distance beyond the cavity. This would require an additional range shifter in front of the target, because the requested energy is rather low and cannot be provided directly by the proton therapy system. Therefore, we added 40 mm PMMA at the front face of the target. Irradiations in the head bear the risk of particle under-range due to cavity filling, which might happen due to swelling or mucus-filling concomitant to a cold. This effect was investigated by replacing the 10 mm air cavity with PMMA. For lung irradiation, we assembled 20 mm adipose, 50 mm muscle, 30 mm lung, and another 20 mm muscle tissue representing the tumour, followed by 30 mm lung, and, finally, we filled with muscle. The beam is intended to stop in the tumour, i.e. the part with 20 mm muscle-equivalent tissue. A problem in lung irradiation aside from the potential motion of the tumour is the need for the beam to cross the rib cage. Due to slight patient mispositioning or respiratory motion, a pencil beam might go through a rib, although it was planned to pass in between them through soft tissue or vice versa. To mimic such an accidentally irradiated rib, we replaced the 50 mm muscle part in the beginning of the target, in accordance with the present patient CT, with 40 mm muscle followed by 10 mm bone tissue. A photograph of both targets, head and lung, is shown in figure 2. The composition of the tissue-equivalent Gammex materials is given in table 1. 2.2. Irradiation
Experiments were performed at the Proton Therapy Center Czech in Prague, Czech Republic. The target was positioned on the patient table with the inhomogeneities along the beam axis and the expected Bragg peak position at the isocenter. Irradiation was performed with monoenergetic proton pencil beams along the target axis and without scanning. Different range modification scenarios were investigated. In the case of the head target, the beam energy was chosen so that the beam stops 5 mm (137.5 MeV) and 15 mm (144 MeV) beyond the cavity. In both cases, we imitated the effect of range shifts (which might have different reasons such as mispositioning or density changes) by changing the beam energy in 10 steps of 1 MeV 4852
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 3. Left: detector of the slit camera. Right: full slit camera with trolley and knife-
edge-shaped slit collimator.
each to higher and lower ranges, resulting in 21 different irradiations for each case. The same was done with the filled cavity target. In case of the lung irradiation, we chose the beam to stop in the middle of the tumour, i.e. beam energy was 110 MeV. To study the effect of accidentally hitting a rib, we irradiated the rib target with the same energy. For both targets, lung with and without rib, we also varied the beam energy from 100 MeV to 120 MeV in 1 MeV steps for investigation of range deviation effects. In each irradiation scenario, approximately 2.6 · 1011 protons were delivered. 2.3. Camera system
The detection of prompt gamma radiation was performed with a knife-edge-shaped slit camera, as described in detail in Perali et al (2014). The detector consists of two rows of 20 LYSO slabs, which are read-out by silicon photomultipliers (SiPM). The detector is mounted onto a trolley, which also holds the tungsten knife-edge slit collimator, figure 3. The distance between beam axis and collimator was set to 25 cm and the distance between detector and collimator was set to 20 cm, resulting in a magnification factor of 0.8. The two rows of 20 LYSO slabs with a crystal width of 4 mm each thereby produced a field of view (FOV) of 10 cm along the beam axis. The center of this FOV corresponding to the slit aperture of the collimator was aligned at the expected range of the impinging protons. The distance between the collimator front face and the target surface is 15.5 cm. A schematic overview on the experimental setup is shown in figure 4. Simulations, measurements, and calculations related to these experiments indicate that the camera exhibits a point spread function between 15 mm and 30 mm full width at half maximum (FWHM), depending on the geometrical setup and the size of the phantom. In the configuration used here, FWHM of the point spread function is approximately 20 mm. Before starting the experiments, energy calibration was performed as proposed by Perali (2014): first with 137Cs (662 keV) and 60Co (1173 keV and 1332 keV), then with the 1.37 MeV and 2.75 MeV gamma lines going along with the decay of 24Na (which was obtained from irradiation of an aluminum sample with protons according to 27Al(p,n + 3p)24Na), and finally 4853
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 4. Schematic overview of the experimental setup. The drawing is not true to
scale. A few prompt gamma photons are shown as green lines.
with the double escape peaks from excited 12C and 16O (3.42 MeV and 5.11 MeV) during proton irradiation of water. These energy calibration values have been fitted with an exponential characteristic to take into account the saturation of the SiPMs microcells for the high amount of light that arises at high energies. On the basis of this calibration, the lower and upper energy thresholds for the measurement of prompt gamma detection profiles were set to 3 MeV and 6 MeV, respectively. Only prompt gammas detected in between these thresholds are scored, because simulations by Smeets et al (2012) and measurements by Perali et al (2014) show that the prompt gamma rays in this energy window exhibit the best correlation with the beam penetration depth for the present camera design. Lower and higher energy gammas suffer from more contamination by uncorrelated events and are therefore discarded. 2.4. Simulation
The performed experiments were simulated with an analytical implementation based on a library of profiles pre-computed by the Monte Carlo PENH algorithm (Sterpin et al 2013). By means of a continuous slowing-down approximation (CSDA), the range of the protons and, consequently, the expected range deviation are obtained. Furthermore, to allow for a better interpretation of the data, the resulting prompt gamma emission as well as the expected detected profile have been simulated. As shown previously in measurements and full Monte Carlo simulation (Smeets et al 2012), the uncorrelated background signal detected by the camera as a result of neutron emissions and their secondaries appears to have an overall rather flat contribution to the detection profile. Thus, it is not simulated in this investigation. The materials used for the experiment, i.e. Gammex tissue-equivalent materials, were modelled very carefully in terms of their elemental composition. However, uncertainties in the ionization potential of those materials remain and affect simulations in terms of quantitative aspects. Therefore, the resulting particle ranges have an accuracy not better than 1 mm. A direct, quantitative comparison between measurement and simulation is beyond the scope of this article and is presented elsewhere (Sterpin et al 2015). 4854
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 5. Procedure of distal edge detection for two different measured profiles (left
and middle) with the help of their first and second derivatives. The dash–dotted vertical lines indicate the detected distal edge determined as described in (ii) in section 2.5. Right: for range deviation estimation, the two parts detected as in the left and middle parts of the figure are normalized and shifted against each other to minimize the area between them. This area is enclosed from the two curves and the dashed lines.
2.5. Data analysis
With the slit camera described one obtains one-dimensional prompt gamma profiles. The measurements consist of 20 single values, one for each 4 mm wide detector bin (i.e. the sum of two crystal slabs: the one from the lower and the one from the upper row). For analysis, a linear interpolation between these values is performed so that one point per millimeter in the target is obtained. A Gaussian function with 20 mm FWHM is applied to smooth the data and reduce statistical noise. Because the width of the smoothing function is not larger than the camera’s point spread function, it is expected that only patterns with spatial frequencies due to statistical fluctuations are reduced, providing no information regarding the real emission profiles. Evaluation of the range shift relies on the investigation of the distal edge of the measured profile. For the prompt gamma measurements presented here, different methods of distal edge detection were investigated. The best procedure here is found to be the following (figure 5): (i) Normalization and background subtraction: all profiles are scaled so that the mean of the three most proximal data points is one and the mean of the three most distal points is zero. (ii) Distal edge definition: it is defined as the most distal part of the profile, where data points are larger than a certain threshold (here, we chose 0.1 to suppress statistical background fluctuations), the profile is decreasing (first derivative 10 mm) in this target the accuracy of range shift estimation suffers slightly (errors up to 2 mm). This can be explained by the applied normalization to the profiles described here and originates from the increasingly different shape of the profiles compared to the reference. In case of the head target with empty cavity, the achievable accuracy depends very much on the distance between the cavity and the planned beam penetration depth. Two different scenarios have been studied. In one case, the beam is intended to stop 5 mm beyond the cavity (137.5 MeV proton energy), and in the other one the beam is to stop 15 mm beyond the cavity (144 MeV proton energy). Figure 8 shows simulations of emitted and detected gammas together with the HU distribution of the target for the empty cavity case. It can be seen that 4859
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
at 137.5 MeV, although the pencil beam stepped across the cavity, there is no corresponding information in the detection profile. Prompt gamma emissions reach their maximum at 25 MeV proton energy, which corresponds to approximately 6 mm residual proton range in water, and then starts declining very steeply to zero at approximately 6 MeV (i.e. 0.5 mm in water). Thus, there is basically no prompt gamma signal in the last 3 mm of the proton path. When the beam stops 5 mm beyond a 10 mm cavity, only approximately 2 mm of tissue beyond the cavity deliver a prompt gamma signal, whereas virtually no signal comes from the cavity itself. This would be measurable with a perfect camera. However, with our real camera the emission profile can be seen as multiplied by a factor of 4 · 10−4 for the efficiency, a point spread function of 20 mm FWHM, and a 1 : 1 signal-to-noise ratio. Therefore, the profile of 137.5 MeV contains almost no information on the cavity that was crossed by the pencil beam. In contrast, the 144 MeV profile, i.e. when the beam stops 15 mm beyond the cavity, shows evidence that the pencil beam crossed the cavity. In this case, overshoots can be detected with excellent accuracy as in the filled cavity cases. Undershoots, however, are recognized by the camera with even shorter range than their actual value, which is related to the missing signal from the cavity. Those points are detected with the actual range shift minus the cavity size and are therefore on a straight line parallel to and below the one representing the equality between estimated and expected shifts (figure 7). In the case of the beam stopping 5 mm beyond the cavity (137.5 MeV), undershoots are underestimated and overshoots are overestimated. Most data points in figure 7 (left) are on a straight line that follows the equation ‘detected range shift = actual range shift + cavity thickness’, i.e. parallel above the line representing equality between estimated and actual shifts. That means the size of the cavity is inherently included for the detected range shift because the reference profile contains almost no information that the cavity was actually stepped across by the pencil beam. Only in a very small region around the planned beam penetration depth (±2 MeV) can a sufficiently high accuracy be realized, because here the detection profile is also missing evidence that the cavity was stepped across. In the investigated lung target, a rather high accuracy for range deviation detection is found as long as the beam stop is in the tumour. This is indicated with black symbols in the respective part of figure 7. As soon as the beam range does not reach the tumour volume anymore, accuracy declines enormously (blue symbols in the right part of figure 7). The correlation between prompt gamma production and proton range seems to be lost in the lung tissue due to its low density compared to the surrounding tissue. 3.1.2. Detectability. As discussed, differences between detected prompt gamma shift and actual existing shift in particle range can occur depending on the composition of the target. The question to be answered now is, to what extent can the shift be detected even though the accuracy of the range deviation estimation can be weak in the presence of large density gradients? To answer this question, ROC curves are used. In figure 9 the histograms of detected shifts for the empty cavity head irradiation with 144 MeV and a shift of −5 MeV and +5 MeV, respectively, are shown for different numbers of incident protons. This energy shift results in a range shift of −8 mm and +8 mm, respectively. For the full statistic profiles (2.6 · 1011 protons) measured in the experiment, a good accuracy of less than 2 mm deviation from the real particle range shift was achieved for irradiation with larger energy (overshoot). However, for irradiation with lower energy (undershoot) the accuracy of shift estimation was rather poor, with a deviation of up to 10 mm from the applied shift due to the increasing influence of the cavity, as can be seen from figure 7 and as was discussed in the previous section. Nevertheless, in figure 9 it can be seen that in both cases, positive as well as negative shifts, the histograms can be separated quite well. This leads to comparatively high and very similar sensitivity and specificity 4860
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
(100 %, 100 %)
(100 %, 100 %) (98.4 %, 99.5 %)
(95.9 %, 96.3 %) (76.6 %, 67.0%)
(74.8 %, 75.0 %)
Figure 9. Top: histograms of detected shifts for the head target (empty cavity) with 144
MeV protons (green) and an energy shift (red) of −5 MeV (left, results in −8 mm range shift) and +5 MeV (right, results in +8 mm range shift), respectively. The distributions are shown for different numbers of incident protons. Bottom: ROC curves for these histograms. The inserted points show the points of maximized Youden index with sensitivity and specificity given in brackets. Note that the ROC curve for 109 protons is hidden below the one for 1010 protons.
for −5 MeV and +5 MeV shift, although the accuracy in shift estimation is very different. For irradiation with 107 protons, a sensitivity of approximately 75% and 77% and a specificity of approximately 75% and 67% (−5 MeV and +5 MeV, resp.) can be achieved for that special anatomical case. For 108 protons, more than 95% sensitivity and specificity can already be reached. That means that the detectability of the shift is equal for both cases, i.e. the shift can be detected with the same probability; however, the accuracy of the estimated shift is very different. The results of all energy shifts in terms of sensitivity and specificity are summarized in figure 10. As expected, for minor shifts in energy and, thus, in range, sensitivity and specificity are comparatively low, because the histograms of detected shifts are difficult to separate for the shifted and non-shifted case. For an increasing number of incident protons, an improvement can be seen. For the shift detection in the lung irradiation exemplary histograms for −2 MeV and +2 MeV shift are shown in figure 11. This shift in energy results in a range shift of −3 mm and +3 mm, respectively. It can be seen that for a lower number of protons, both histograms become difficult to distinguish. The results on the accuracy presented show that good accuracy of less than 2 mm can be achieved when statistics are very high. However, the results of the 4861
M Priegnitz et al
Phys. Med. Biol. 60 (2015) 4849
Figure 10. Sensitivity and specificity for the detection of range shifts in the head target
(empty cavity) due to variations in the proton energy around the nominal energy of 144 MeV.
analysis of ROC curves show that sensitivity and specificity decrease severely when lowering prompt gamma statistics, i.e. decreasing the number of incident protons. Results concerning sensitivity and specificity for all investigated shifts in energy are shown in figure 12. For irradiation with 108 protons or more, rather good sensitivity and specificity of more than 70% can be achieved in most cases. Although the accuracy in the range shift estimation is not satisfying for large shifts due to the missing correlation between proton range and gamma emission in low-density tissue (figure 7), the ROC curves show that the existence of the shift can be generally detected. Thus, further examinations on the source of the range deviations related with their potential consequences on the treatment delivery can be initiated. 3.2. Range deviation due to cavity filling or a rib 3.2.1. Accuracy. Another question to address is the accuracy and detectability of range devia-
tions due to cavity filling in the head target or accidental irradiation of a rib in lung irradiation. Therefore, the measured profiles of the head target with empty cavity and the lung target without rib have been taken as references. For range deviation analysis, the profile from the target with filled cavity and the target with rib, respectively, with the same proton energy have been used. The results are summarized in figure 13. Here, the detected range deviations are plotted as symbols for each proton energy. The expected range shift from simulation is given as a grey solid line. In the case of head irradiation (figure 13 left), data from two different setups are available. In the 137.5 MeV setup, the slit of the collimator was positioned at the range of the 137.5 MeV protons, and then the proton energy was changed ± 10 MeV in 1 MeV steps. The other setup is the 144 MeV setup with the slit of the collimator at the range of the 144 MeV protons. Thus, there is an overlap in proton energy from both measurement series (134 MeV–147 MeV). However, detected shifts do not coincide completely in that overlap due to a different target region within the FOV of the camera in both setups. For the low energies (
