Radiotherapy and Oncology, 24 (1992) 261-271

© 1992 Elsevier Science Publishers B.V. All rights reserved. 0167-8140/92/$05.00

261

RADION 01003

Maximizing setup accuracy using portal images as applied to a conformal boost technique for prostatic cancer Jurrien Bijhold, J o o s g. Lebesque, Augustinus A. M. H a r t and R o n E. Vijlbrief Department of Radiotherapy, The Netherlands Cancer Institute (Antoni van Leeuwenhoek Huis), Amsterdam, The Netherlands

(Received 8 July 1991, revision received 3 March 1992, accepted 6 March 1992)

Key words: Radiation therapy; Radiation field alignment; Patient setup verification; Decision rules; Quality assurance; Portal imaging;

Conformal field therapy; Prostatic cancer

Summary A design procedure of a patient setup verification protocol based upon frequent digital acquisition of portal images is demonstrated with an application for conformal prostatic boost fields. The protocol aims at the elimination of large systematic deviations in the patient setup and includes decision rules which indicate when correction of the patient setup is needed. The decision rules were derived from the results of a theoretical and quantitative analysis of patient setup variations measured in three pelvic fields (one anterior-posterior and two lateral fields) of 105 fractions for nine patients. Deviations in the patient positioning, derived from one field, were quantified as two-dimensional (2-D) displacement vectors in the plane perpendicular to the beam axis by alignment of anatomical features in the portal and the simulator image. The magnitude of the overall setup variations along the anterior-posterior, superior-inferior and lateral directions varied between 2.6 and 3 mm (1 S.D.). Inter- and intratreatment variations could be separated, both having equal magnitudes of 1.7 to 2.2 mm (1 S.D.). In addition, intra-treatment variations appeared to be predictable which was a prerequisite for the development of the decision rules. The 2-D setup deviations, measured in the three fields of one fraction were strongly correlated and a 3-D displacement vector was calculated. Utilization of this 3-D vector in a setup verification protocol may lead to an early detection of systematic setup deviations.

Introduction

High accuracy of the patient setup for radiation therapy is imperative when conformal treatment fields with tight tumor margins are used. Especially for the boost treatment of prostatic cancer, conformal static field therapy was developed in our institution [ 10] as well as in other centers [ 11,15,17 ]. A straightforward strategy to achieve a higher overall accuracy for such treatments is to reduce systematic deviations in the patient setup by running a setup verification protocol during the course of treatment. Such a protocol could be based on regular inspec-

tion of portal films. Studies have been reported on the quantification of setup deviations for all common fields using portal films or images [4,12,14,15]. These deviations were quantified as differences in measured distances between anatomical features and field edges in a portal and simulator film. The methods to determine these deviations are rather time consuming and consequently not very suitable in a setup verification protocol that requires frequent measurements of patient setup deviations. In order to keep the extra workload to an acceptable level, digital acquisition of portal images and a simple method for the measurement of setup deviations is de-

Address for correspondence." J. V. Lebesque, Department of Radiotherapy, The Netherlands Cancer Institute (Antoni van Leeuwenhoek Huis),

Plesmanlaan 121, 1066 CX Amsterdam, The Netherlands.

262 sired. A digital portal imaging device [ 13,19] and a fast method for the measurement of setup deviations [2], which does not require distance measurements, were developed in our institution. In this study we only deal with deviations of the patient positioning, since the collimator positions can be controlled by narrow tolerance settings of modern accelerators. A patient setup verification protocol should indicate when setup deviations have to be measured, and when a setup correction is needed. In some studies [3,4,14] setup deviations were split up into systematic and random contributions by analyzing the measurements in series of portal images. The systematic contributions in these studies were due to differences in the patient setup for simulation and actual treatment and random contributions were due to day-to-day variations in the patient setup during the course of treatment. Because the number of corrections should be as small as possible in clinical practice, a decision to correct the patient setup should be based on evidence for a real systematic deviation of the patient setup which can indeed be corrected for. The last condition means that the origin of the systematic deviation should be understood. The aim of this study was to test our concepts for the design of such a verification protocol. To understand which contributions to the measured patient displacement vector exist and which of these contributions can be reduced by setup corrections, a model of these contributions and their variations is devised and tested with clinical data. Subsequently, decision rules are derived that indicate when the patient setup has to be corrected. We choose to perform this study for conformal prostatic boost fields because the high dose treatment for prostatic cancer demands a high accuracy.

0.66) in the three fields resulted in the desired weight factor for the large-field treatment. Using the Beam'sEye-View (BEV) option of our planning system the field shape of the simultaneous boost was designed and cutaway in the transmission plate. The quality of the portal images of the fields with transmission plates was excellent (Fig. 1) and permitting anatomical details outside the boost volume to be distinguished; these details were used for an accurate measurement of the anatomy position in the image. Treatment planning, simulation and actual radiation treatment Treatment planning started with making a CT-scan of the pelvic region of the patient in treatment position (supine) on a fiat table top. The CT-slice thickness and spacing was 5 mm. Using alignment lasers, a tentative isocenter was marked and tattooed on the patients' skin surface, and radio-opaque catheters were placed on these skin marks. This CT-information was entered into our threedimensional planning system (Scandiplan). The outlines of normal structures like bladder and rectum and the contours of the immobile biological target volume (prostate and seminal vesicles) were marked in every slice. A margin of 7 mm to account for tumor motion within the patient and patient setup variations, was

Materials and methods

The simultaneous boost technique for prostatic cancer In our institution the conventional treatment schedule for the irradiation of stage Tzb, T3 and T 4 prostatic cancer consists of a large-field treatment for the first 20 fractions of 2 Gy and a subsequent boost irradiation of 15 fractions of 2 Gy, resulting in 35 fractions of 2 Gy to the boost target volume. Both for the large field and for the boost field treatment a three field technique is used with one anterior-posterior field and two wedged lateral fields. With the simultaneous boost technique [ 10] for prostatic cancer the large-field treatment is delivered simultaneously (i.e. within the same few minutes) with the boost dose for all 35 fractions with appropriate weight factors. In practice, Cerrobend transmission plates with a thickness of 9 mm (transmission factor of

Fig. 1. Portal images of the AP field (a,b) and the right lateral field (c,d) before and after image processing with a Laplacian filter. The dark regions in the unprocessed images correspond to the simultaneous conformal boost fields.

263 added to these contours to form the boost target volume, as defined in ICRU Report-29 [6]. Beam's-Eye-View projections of this boost target volume were calculated for the three beam directions together with boost field edges, which conformed to the projected boost target volume with a 3 mm margin. This additional margin for the field edges was necessary to ensure that the target volume was fully encompassed by the 95~o isodose (the 100~o being defined in the isocenter). Utilizing these three projections the transmission plates with the appropriate BEV-designed cutaways could be fabricated. Finally, Digital Reconstructed Radiographs (DRRs), were computed for the three fields, together with the outlines of large and boost field edges. During the treatment simulation the patient setup was corrected until the simulator image fitted to the DRR for every field. New skin marks for each field center were applied to make the corrected setup reproducible during treatment. Subsequently simulator films were exposed, which served as reference images to be compared with portal images during the actual treatment. The films were exposed twice, with and without the transmission plates, to mark the outlines of the large fields and boost fields on the same film. Only two patients were fixed in a cast during the complete treatment. This casting system* consists of an evacuated airtight plastic bag with a content of polysterol microspheres [7]. The other patients were placed in a supine position on a 2 cm thick foam mattress. The setup was verified with portal images of the three treatment fields about twice a week. The field setup of one of the patients had been corrected after the third setup verification. The data of this patient were omitted in the statistical analysis of the remaining nine patients in the study. The setup verification protocol which was developed is demonstrated for this patient, retrospectively. The actual delivered dose in the isocenter was carefully checked each week by high-accuracy in vivo dosimetry [ 5 ].

Portal image acquisition and processing The electronic portal imaging device we used, was developed in our institution for on-line acquisition and presentation of portal images during irradiation [13,18,19]. The device is based on a 256 x 256 matrix of liquid-filled ionization chambers with a sensitive area of 32 x 32 cm.

Bony structures both in the high and low dose regions of the portal images (Fig. la,b) were visualized by applying image processing based on edge enhancement with a Laplacian filter (Fig. lc,d). Sharp edges in the original images are now situated on the sharp black to white transitions in the processed images. A radiation field edge detection algorithm [ 1] was applied to detect field edges in the portal images before they were processed.

Measurement of setup deviations The method used for the measurement of patient setup deviations was developed in our institution and is based on the alignment of field edges and anatomical features in the simulator image and the portal image [2]. Since the patients were treated on a Philips SL-25 accelerator with narrow tolerance settings for its built-in verification of collimator position, these setup deviations concern only the positioning of the patient. These deviations are quantified by a translation vector (the 2-D displacement vector) and a rotation angle. The position of the simultaneous boost fields with respect to the outlines of the large fields is determined by the mechanical stability of the tray holder and the reproducibility of the position of the tray with the transmission plates in this tray holder. Utilizing the same method as used for measuring patient setup deviations, deviations of the position of the boost fields were less than 0.7 mm.

Theoretical model of setup variations The displacement vector measured for a patient, p, and fraction, f, of a treatment series is referred to a s mpf. This vector is assumed to be the sum of the actual displacement vector and a measurement error vector, and each of these two vectors is assumed to have a systematic and a random contribution:

mpi= Ap + bpi + Ep + epi

(1)

The vector Ap is the systematic displacement of patient p during the whole treatment and the vector bpf is the random deviation in each fraction f. The vector Ep represents the systematic error made during the measurements for patientp. This error is due to different interpretations of the anatomy and/or the field edges in the simulator and the portal image. Random errors, given by the vector ~pf, are caused by imperfect alignment of the simulator graph on the portal image.

* Vac-Fix system, Marshall Med., Eskemosegyde 18, DK 5600 Faaborg, Denmark.

264 Increased setup accuracy by setup corrections can only be obtained by reducing the systematic deviation Ap, because reducing bpfwould require, above all, better patient positioning techniques. This argumentation suggests that we should treat the systematic (Mp) and the random (gpf) contributions to mpf separately:

mpf= Mp + #pf,

Mp = Ap + Ep #pf= apf + epf

(2)

If all these vectors are distributed normally and independently their distribution functions can be represented with 2 x 2 covariance matrices. The covariance matrices for the systematic and random contributions are referred to as E and ap, representing inter- and intra-patient variances, respectively. A necessary condition for any decision rule for setup corrections is prior knowledge of the intra-patient variances in ap. Our basic assumption is that the intrapatient variances are equal for all patients and thus predictable. This predicting covariance matrix is referred to as a. The covariance matrix for the distribution of a vector mpf is now given by E + a. If no errors are made in the transfer of setup data from the simulator session to the first treatment session, an estimation of the values in E is obtained from the following arguments. Patients are positioned on the treatment table using laser beam alignment with marks on the skin. The patient position is measured, however, using the features of the pelvic bone in both the simulator and the portal image. For obvious reasons the skin marks are not in an accurately reproducible position relative to the pelvic bone. Distribution functions which relate skin mark positions and pelvic bone positions could be determined for simulation and treatment sessions. When the simulation and treatment procedures are carried out correctly, these distributions are likely to be equal. The skin marks which are set during the treatment simulation are just one sample of this distribution. When these marks are used for patient positioning during a treatment session, the resulting position of the pelvic bone is just another sample of the same distribution. The vector Ap, representing the systematic difference between the pelvic bone position during simulation and the average position during treatment, should therefore be distributed like bpf. In case the variances of Ep and Spy are small or just equal to each other, values in Z should be equal to values in a. These assumptions, i.e. ap = o-for all p and Z = o', are tested from an analysis of the measurements performed in this study. The matrices a and Z are estimated using these assumptions.

The three-dimensional (3-D) patients displacement vector The aim of combining measurements of the three orthogonal fields is to get a more accurate estimation of the actual 3-D patient position because random measurement errors will reduce due to the averaging process. Such a kind of averaging makes only sense when the patient position does not change when the field setup is changed from one field to the next field. The displacements measured in the anterior-posterior (AP), right (R) and left (L) lateral field (from the patient's point of view) are referred to as mAp , mn and mL and the 3-D patient displacement vector as ms_D. All these vectors are expressed in gantry coordinates for zero gantry angle. In this coordinate system the x-axis is pointed in the left lateral direction, the y-axis in the inferior-superior direction and the z-axis in the posterior-anterior direction. The components of mAe, mn and mL are referred to as XAe, YAP,ZR, In, ZL and YL. The 3-D vector is given by (XA?, Y,~, Zav) in which Yav = (YAP + Yn + YL)/3 and Zav = (ZR + ZL)/2. The distribution of this 3-D vector is also assumed to be normal and is represented with a 3 × 3 covariance matrix.

Decision rules for setup corrections Having estimated acceptable values for systematic (E) and random (a) setup deviations (inter- and intratreatment variations, respectively), an elliptical confidence region or volume can be computed from the matrix (E + a) at a specified confidence level, e.g. 95 or 99%. This confidence region is expected to confine 95 or 99% of measured vectors mpf. A decision rule may state that the patient setup is to be corrected, when a vector mpf is measured which is outside this region. When the correction has been carried out, new skin marks are applied on the patient to make the correction effective for the consecutive treatment sessions. The setup verification protocol with decision rules for setup corrections is aimed at safeguarding tolerance levels for systematic setup variations. The random contribution to a setup deviation may lead to erroneous decisions; corrections may be unnecessary or cancelled when needed [9]. If we use a decision rule after each measurement with a 95 % confidence level and E is about equal to a, the average number or erroneous decisions will be about half of 5 % of the number of measurements. The probability of one erroneous decision during a full treatment with 12 measurements will be about 26%. This probability reduces to about 6%, if the 99% confidence level is used. When a decision rule for a setup correction is based upon the average vector computed from the vectors

265 which have been measured since the start of a treatment or since the last correction performed, a different picture emerges, since after a number of measurements a smaller uncertainty is obtained in the values of Mp, as estimated by this average vector. The covariance matrix for the distribution of this average vector is given by Z + a/n. The computed confidence region for the average vector becomes smaller for larger n expressing the increasing certainty on the value of Alp. Consequently, as n increases the probability of an erroneous decision reduces and a smaller confidence level for the decision rule can be applied.

Setup verification protocol The verification protocol is based upon the current clinical practice of taking one portal image at every third fraction and aims at detecting serious systematic deviations. We suppose that our assumption of about equal systematic and random deviations (i.e. Z is about equal to a) holds true. We start with a measurement at the first fraction to detect errors in the transfer of setup data from simulator to treatment machine. The decision rule is applied at the 99~o confidence level, if a decision has to be made after one measurement. If the deviation is between the 95 and 99~o confidence level, a measurement is performed the next day. As the number of measurements (n) increases, the confidence level for the decision rule is lowered. For n = 2 and 3 we take the 95~o level, whereas for n f>4 we use the 67~o level. Consequent application of this protocol means that at least 33 ~, of the patients have a setup correction during their treatment. Some of these patients, however, will have two or more setup corrections in case the first correction appears not to have been effective. In order to make corrections effective, they should be carried out accurately. Accurate in this respect means that the variations due to alignment of the laser beams to the skin marks should be smaller than the variations due to the relative positions of the skin marks and the pelvic bone.

combination of the three images (i.e. the 3-D analysis). The third section illustrates the use of the decision rule using the data from the tenth patient whose setup had been corrected during the treatment.

Field setup deviations The results of measurements for one patient are shown as an example (Fig. 2). The 2-D distribution of the central black dots in this figure (representing the measured patient displacement vectors) is described by a 2-D Gaussian function. The average measured dot position is an estimation of the actual systematic setup deviation for this patient (Alp). The measured variances along the horizontal and vertical axis and their covariance make up the sample covariance matrix which is an estimation of O-p.The values for the average rotation of the portal fields and its variance complete the description of the patient setup variations for this particular patient. Relatively small variances were expected for the two patients that were treated in a cast. However, it appeared that relatively large variances were found for one of these two patients. The 2-D distribution of the 105 patient displacement vectors mAe, mn and mE measured in the portal images

Results The results are presented in three sections. The first section gives an overview of the data obtained from the measurements with the 315 portal and 27 simulator images for nine patients. In the second section these data are analyzed statistically and the assumption of the theoretical model of setup variations are tested. Three separate analyses were performed on data from AP field images, on both lateral field images and on the

Fig. 2. Simulator image of the AP field with overlays of its graphics presentation and the field edges in the portal images which have been transformed to correct for the measured patient displacements. The black dots close to the beam axis are the centers of gravity of the portal field edges representing the patient displacement vectors mAe (the position of the beam axis is the origin of these vectors).

266 TABLE I

of the nine patients can be represented in a graphical way by plotting the endpoints of these vectors (Fig. 3ac). The 3-D distribution of the 3-D displacement vectors m3_D can only be visualized by showing the projection on one or more planes (Fig. 3d: projection on the y,z-plane). The large dots indicate the average position of a patient (representing Mp). The sample covariance matrix that is computed from the vectors Mp is an estimation of Z. The ellipse shaped regions in Fig. 3 show the 95 ~o confidence regions of the sample distributions. The length and direction of their principal axes were computed from the sample covariance matrix of all small dots in Fig. 3. Their positions were computed from the overall average dot positions (which are estimations of the overall systematic displacements, M). The random or day-to-day variations of all patients are visualized in the scatter plots of Fig. 4, which were created from the original dots in Fig. 3 by subtracting the corresponding vectors Mp. The sample covariance matrix of this newly created distribution is an estimation of the predicting covariance matrix a. Comparison of Figs. 3 and 4 suggests that the apparent ellipticity of the confidence regions in Fig. 3 is caused by the distribution of the vectors Mp. 1B

The average values of the magnitude of the average shift (A-) and its standard deviation (S.D.) per patient, in the lateral (x), inferiorsuperior (y) and the posterior-anterior direction (z). AP

1B

Shifts

A

S.D.

x y z

1.2 2.0

1.7 1.6

glP

1B

A

S.D.

A

S.D.

1.5 1.2

1.7 2.3

2.0 2.4

1.9 2.2

R

1" ¥

X

~

18

L

-18 1B

1'

1'

Y

Y

-1B

L

To facilitate in the discussion the comparison of random and systematic setup variations with the results of Rabinowitz et al. [14], Huizenga et al. [4] and Griffiths et al. [3], another set of statistics was calculated (Table I). The components of the displacements vectors map , m n and mL are treated separately as the lateral, inferior-superior and anterior-posterior shifts. The magnitude of the average shift and its standard deviation was computed per patient and per field. The average values of these two numbers for all patients (Table I) represent the average absolute systematic shift (A) and the average random shift (S.D.). The rotation measured in the portal images of these

t ¥

-18

R

Z

~

18

-1B

Z

•*

18

~

18

3D

Z

Fig. 3. Scatter plots showing the 2-D displacement vectors by small dots in the AP (a) and lateral fields (b,c) and the 3-D displacement vectors projected on the y,z plane (d). The large dots show the average displacement vectors for every patient. The grey regions in every plot show the 95 % confidence regions. The displacements are given in mm along the lateral (x), inferior-superior (y) and the posterior-anterior (z) directions.

267 18

18

AP

R b

i i i ~ ! i i ii!

,=~iiliii

t Y

=ili~iH~,.)£~iiiii~!i==

i

-18

X

~

18

L

10

I

-10 18

t

t

!

:

i

i

Z

1B

3D

r

.......

Y

I

-1B

Z

~

1B

I

I

-18

Z

">

I

18

Fig. 4. Scatter plots showing the 2-D displacement vectors (a,b,c) and the 3-D vectors projected on the y,z plane (d) after subtraction o f the corresponding average displacement vectors. The grey regions in every plot show the 95% confidence regions. The displacements are given in mm along the lateral (x), inferior-superior (y) and the posterior-anterior (z) directions.

patients appeared to be smaller than 1 degree in 90~o and smaller than 2 degrees in 95 To of the images. The largest rotation measured was 3.3 degrees. Because the effect of such rotations on the dose delivery are small compared with the effects of patient displacement, they have not been analyzed any further. Analysis of variances

The data in Figs. 3 and 4 for the vectors mAp , mR, mL and m3_D can be represented by an overall average displacement M, a sample covariance matrix Z, computed from the 9 vectors Mp, and 9 sample covariance matrices O-p.A statistical analysis was performed to estimate the "real" values in M, Z and O'p for each of these vectors, i.e. the values that would be computed from an infinite number of measurements per patient and an infinite number of patients. Common estimations of these values were computed for the two vectors mR and me. All estimations were performed using random effect variance analysis with restricted maximum likelihood, the validity of assumptions made during these estimations was tested using the likelihood ratio test with a correction for small sample size according to Box [8].

Firstly, our assumption that O-p= ~ for all patients p was tested for the vectors map , mR, mL and ms_D. The usefulness of common estimations of M, Z and afor mR and m/_ was justified by testing the assumptions that their corresponding covariance matrices are equal. The usefulness of the 3-D vector m3_D was demonstrated by testing the correlations between YAe, YR and YL, and the correlations between ZR and ZL. Then, two sets of estimations were computed for each vector; with and without using the assumption that 12 = a. Finally, tests were performed for the assumptions that E = o', that M = 0 and that covariances are zero. The Tables II and III show the results of the tests

T A B L E II Results of the tests of the four assumptions. Yes means there was no evidence for rejection. Assumption

AP

R, L

3-D

1: a = a p 2: Z = a 3: M = 0 4: Covariance terms are zero

yes, if p = 1..9 yes yes

i f p = 1,3,5..9 yes no

i f p = 1,4..9 yes

yes

yes

no

IlO

268 and the estimations, respectively. Note that our basic assumption, the predictability of ap of the vectors map, mR, mL and m3-D, was demonstrated for a majority of the patients. Furthermore, there is an indication of an overall systematic displacement in the lateral fields of all patients, M S 0, and there may be prevalent directions of displacements, since the covariances of the 3-D vector m3_D do not equal zero.

The use of the decision rule for the vector mAe is illustrated in Fig. 5a,b. The ellipses represent the 99~o (largest one), 95 and 67~o (smallest one) confidence regions of the distribution (E + a/n) of average displacements after 1 and 2 measurements (n = 1 and 2). As actual data, measured setup deviations are shown of the patient for which a setup correction was carried out. In the protocol, a setup correction after one measurement is prescribed when the measured deviation is outside the 9 9 ~ confidence region, and a measurement should be performed the next session if the deviation is between the 95 and 99~o confidence level. Since the first measured setup deviation (Fig. 5a) was inside the 95 ~o confidence region, no steps were taken. After the second measurement (Fig. 5b), the protocol demanded correction of the patient setup on the next day, since the average setup deviation is outside the 95 ~o confidence region. The value of the measured average displacement in the y-direction was used for this correction. After the setup correction for this patient, the average setup deviation of the subsequent 8 measurements

Decision rules and setup verification protocol For the application of decision rules we have to accept or reject the assumptions which are listed in Table II. For the vectors map, mR and mL we accepted the assumptions E = a, M = 0 and covariance terms equal to zero. For the vector m3_D we accepted the assumptions E = tr and M = 0 but rejected the assumption of zero covariance terms. The assumption M = 0 was accepted for all vectors because we did not want to accept displacements which are systematic for all patients, whatever its origin.

1B

AP

1B

AP

1' ¥

-18 1B

X

4

18

P.P

-18 18

X

~

1B

~

1B

3D

'!"

t

~t Y

-1B

X

~

1B

-1B

X

Fig. 5. Scatter plots which illustrate the decision rule for correction of the patient setup for an AP field based upon the measured average 2-D and 3-D displacement vector after 1 (a) and 2 (b,d) measurements, and after setup correction and 8 subsequent measurements (c). The small dots show the measured displacements. The large dots show the average displacements. The ellipses in every plot represent the 99% (largest), 95 and 67% (smallest) confidence regions of the predicting distribution of average displacements. The grey regions show the confidence regions which are used for the decision. The displacements are given in mm along the lateral (x), inferior-superior (y) and the posterior-anterior (z) directions.

269 (Fig. 5c) is within the 67 ~o confidence region, illustrating the success of the setup correction. The use of the decision rule with the vector m3_D after two measurements is illustrated in Fig. 5d, which shows a 2-D projection of the 3-D vector-space. The distance between the edge of the 95 % confidence region and the average displacement vector is larger than in Fig. 5b suggesting that the use of this 3-D vector could lead to a setup correction after less measurements, and fractions given. Note that for proper use of the 3-D vector a second projection on one of the two orthogonal planes should also be inspected.

Discussion

Setup deviations and variance analysis The measurements of setup deviations and their analysis showed that intra-treatment variations are equal for most patients. This implies that these variations can be predicted for a new group of patients and that the use of our decision rules make sense. The magnitude of these random variances was estimated to be about 2 mm (1 S.D.). The analysis also showed that there was no reason to reject the assumption that the distributions of inter- and intra-treatment deviations are equal. More patients should be studied for a better estimation of Y~in order to obtain a better proof for equivalence. The magnitude of the combined inter- and intratreatment variances is about 3 mm (1 S.D.). The precision of the method used for the measurements was estimated to be 2 mm from intra-observer variations [2]. Because the magnitude of the measured intratreatment variations was also estimated to be 2 mm, we can conclude that the real intra-treatment variances are probably smaller. Rotations of the patient have not been studied further, because they appeared to be small and have small influence on the total dose delivery. It should be noted, however, that out-of-plane rotations in a field, i.e. the rotation around the y-axis (the longitudinal body axis), cause erroneous results of the measurements when they are larger than 6 degrees [2]. Because the rotations around the z- and the x-axis, measured in AP and lateral images respectively, were all smaller than 4 degrees, it was assumed that the rotations around the y-axis were also smaller. The overall systematic displacement (M) in the lateral fields (see Table I, II) may be caused by a systematic difference in the interpretation of simulator and portal images, or by a systematic difference in the geometry of patient setup during simulation and treat-

ment. The latter explanation will have to be tested with phantom that can be placed accurately and that shows sharp features in a portal image. Such a phantom was described by Meertens et al. [ 12]. A comparison was made with results published by other groups who have analyzed distance measurements in simulator and portal images. Direct comparison with the results in Table I is complicated by the fact that all these groups estimate systematic shifts with less portal images per field per patient. Huizenga et al. [4] showed equivalent values (3 mm) for the average absolute systematic shifts and the average random shifts for a group of 22 head and neck patients with 138 portal images and 55 simulation images. These findings are in agreement with our assumption that Z = a. In the study of Rabinowitz et al. [14], however, systematic shifts were found to predominate over random shifts for various groups of patients. They found for pelvic fields an average absolute systematic shift of 5.6 mm and an average random shift of 2.5 mm (1 S.D.). They estimated the average absolute systematic shifts from a larger number of patients, 23, but only from one portal image per patient. A possible explanation for these high systematic shifts is that gross deviations may have been included in their calculation of average absolute systematic shifts. Our measurements showed that the distributions of the measured displacements are not circular (Fig. 3, Tables I and III). The variations in the inferior-superior direction, along the y-axis, are larger than the variations in the lateral direction, along the x-axis. This difference was mainly caused by the contributions of the systematic displacements (see Tables I and III and compare Figs. 3 and 4). The same observation was made by Griffiths et al. [3] who performed a simulator study on TABLE

III

Estimated values for the overall systematic displacement M and the s q u a r e r o o t o f t h e v a r i a n c e s in Z a n d a in m m a l o n g t h e x, y a n d z axis o f t h e fixed g a n t r y c o o r d i n a t e s y s t e m . Z'= a

AP

R, L

Axis

M

2~

a

M

x

0.0

1.7

1.7

-

y

0.0

2.0

2.0

- 1.5

2.2

z

-

0.0

2.0

2~ve a

AP

2;

3-D a

M

M

Z

a

M

x

0.0

1.2

1.8

-

y

0.0

2.3

1.9

z

-

R - 1.9

Z

a

1.7

1.7

2.2

0.0

1.8

1.8

2.0

- 1.5

2.3

2.3

R, L

Axis

S

0.0

3-D ~

2.3

2.2

L - 3.0

2.3

2.2

0.0

2.3

1.9

M

Z'

a

0.0

1.2

1.8

0.0

2.1

1.8

- 1.1

1.8

2.3

270 the accuracy of different pelvic field setup techniques. They found that both systematic shifts and random day-to-day shifts are larger along the y-axis than along the x-axis. They observed in 72 setups for a group of 20 patients about the same values for the average absolute systematic (2 mm) and random (1.8 mm) shifts as in our study, again in agreement with our assumption E = a. The group of Soften et al. [ 15] did not separate systematic and random day-to-day setup variations. They showed for a comparable group of patients with prostate cancer, a median daily variational shift of 3 mm while the use of a low-density alpha cradle cast resulted in a very low median daily variation of 1 mm. In our study, casts were only used for two patients and therefore no conclusions about the effect of casts on setup accuracy can be made. It is remarkable, however, that we found the lowest accuracy of patient setups for one of these casted patients. Our explanation for this unexpected result is that the skin of this patient, with skin markers, could not move freely enough in the cast during the positioning of the patient before treatment. An improvement for patients in our casts could be to apply markers on the cast instead of the usual skin markers. This might work if the cast is considered, and designed, as a tool which should fix the pelvic bone.

Boost target volume The shape of the boost target volume was conform to the immobile biological target volume (tumor volume) while keeping a distance of 7 mm to account for tumor motion within the patient and for variations in patient position. Motion of the prostate and seminal vesicles within the patient can be estimated to be equal to 4-10 mm from a study of Ten Haken et al. [16]. The measured overall setup variations were equal to 3 mm (1 S.D.). Consequently, the margin of 7 mm is probably too small for our patients to account for all variations.

Decision rules and setup verification protocol The decision rules for a setup correction were based upon confidence levels of the predicting overall distribution (Z + a/n) of (average) setup deviations. The chance of unnecessary corrections depends heavily on the relative contribution of the systematic (Y.) and random deviations (a) in the overall distribution and on the number of measurements (n). When the systematic deviations are relatively large, the decision rule can be based upon a relatively low confidence level, without running the risk of too many needless corrections. On

the other hand, when the contribution of random deviations is relatively high, a setup verification protocol should prescribe online setup verification of each fraction. In this study of prostatic field setups, systematic and random deviations were about equal. The verification protocol for this type of deviations was illustrated for one patient. The demonstration showed that the use of such a protocol results in the early detection of a relatively small systematic setup deviation. Furthermore, the use of 3-D information could lead to an earlier indication of a systematic setup deviation than the conventional use of 2-D information. The measurements performed after the correction for this patient showed indeed a reduction of the systematic deviation. This reduction, however, could have been coincidental. More information about the effectiveness of corrections will be obtained by analyzing the corrections of the next group of patients which are treated using the protocol. Another setup verification protocol will be discussed briefly here. It aims at reducing the systematic deviation Mp as much as possible. Effective correction is then essential. Measurements are performed for every fraction at the start of the treatment until a/n has become small compared with E and the confidence region for the average vector does not change much any more (e.g. n = 5 for our group of patients). The values of Mp are well determined at that time and only then correction is allowed. The treatment may proceed using less frequent measurements.

Conclusions

After theoretical and quantitative analysis of setup deviations for a conformal prostatic boost technique, a patient setup verification protocol was designed, aiming at elimination of large systematic patient setup deviations. The protocol includes decision rules that indicate when correction of the patient setup is needed. Essential for the design of such decision rules was that random or day-to-day variations can be predicted. Furthermore, the relative contributions of systematic and random setup deviations should be known. Otherwise, the use of these decision rules results in too many or too few setup corrections during treatment. The disturbing effect of the random variations on the decision rule could be diminished by varying the action level of the decision rule with the number of prior setup measurements. The effect of random measurement errors could be reduced by combining the 2-D measurements of orthogonal fields of one fraction to a 3-D patient displacement vector.

271

Acknowledgements The authors would like to thank Prof. Dr. J. Strackee, Dr. B. J. Mijnheer and Prof. Dr. H. Bartelink for useful criticism during the preparation of the paper, A. Schouten for helping with the acquisition of the portal

images and M. B. C. Pinkster for assistance during software creation. This work was supported by grants from the Dutch Cancer Society (NKI 87/09) and the Commission of the European Communities (AIM project No. A1024 QUIRT).

References 1 Bijhold, J., Gilhuijs, K. G. A., van Herk, M. and Meertens H. Radiation field edge detection in portal images. Phys. Med. Biol. 36: 1705-1710, 1991. 2 Bijhold, J., van Herk, M., Vijlbrief, R. and Lebesque, J.V. Fast evaluation of patient setup during radiotherapy by aligning features in portal and simulator images. Phys. Med. Biol. 36: 16651679, 1991. 3 Griffiths, S. E., Khoury, G. G. and Eddy, A. Quality control of radiotherapy during pelvic irradiation. Radiother. Oncol. 20: 203-209, 1991. 4 Huizenga, H., Levendag, P. C., De Porre, P. M. Z. R. and Visser, A . G . Accuracy in radiation field alignment in head and neck cancer: a prospective study. Radiother. Oncol. 11: 181-188, 1988. 5 Heukelom, S., Lanson, J. H. and Mijnheer, B.J. In vivo dosimetry during pelvic treatment. Radiother. Oncol., Submitted, 1992. 6 International Commission on Radiation Units and Measurements. Report No. 29: Dose specification for reporting external beam therapy with photon and electrons. Washington, D.C., 1978. 7 Jakobsen, A., Iversen, P., Lindberg Hansen, J. and HjelmHansen, M. A new system for patient fixation in radiotherapy. Radiother. Oncol. 8: 145-151, 1987. 8 Kendall, M. G. and Stuart, A. In: The advanced theory of statistics. Design and Analysis, and Time-Series, Vol. 3; 3nd edn, pp. 265-269. Editors: M . G . Kendall and A. Stewart. Griffin, London, 1968. 9 Lebesque, J. V., Bel A., Bijhold J. and Hart A. A . M . Detection of systematic patient setup errors by portal film analysis. Radiather. Oncol. 23: 198, 1992. 10 Lebesque, J.V. and Keus, R. The simultaneous boost technique; the concept of relative normalized total dose. Radiother. Oncol. 22: 45-55, 1991.

11 Low, N.N., Vijayakumar, S., Rosenberg, I., Rubin, S., Virudachalam, R., Spelbring, D. R. and Chen, G. T.Y. Beam's eye view based prostate treatment planning: is it useful? Int. J. Radiat. Oncol. Biol. Phys. 19: 759-768, 1990. 12 Meertens, H., Bijhold, J. and Strackee, J. A method for the measurement of field placement errors in digital portal images. Phys. Med. Biol. 35: 299-323, 1990. 13 Meertens, H., van Herk, M., Bijhold, J. and Bartelink, H. First clinical experience with a newly developed electronic portal imaging device. Int. J. Radiat. Oncol. Biol. Phys. 18: 1173-1181, 1990. 14 Rabinowitz, I., Broomberg, J., Goitein, M., McCarthy, K. and Leong, J. Accuracy of radiation field alignment in clinical practice. Int. J. Radiat. Oncol. Biol. Phys. 11: 1857-1867, 1985. 15 Soften, E. M., Hanks, G. E., Hwang, C. C. and Chu, J. C. H. Conformal static field therapy for low volume low grade prostate cancer with rigid immobilization. Int. J. Radiat. Oncol. Biol. Phys. 20: 141-146, 1991. 16 Ten Haken, R. K., Forman, J. D., Heimburger, D. K., Gerhardsson, A., McShan, D. L., Perez-Tamayo, C., Schoeppel, S. L. and Lichter, A.S. Treatment planning issues related to prostate movement in response to differential filling of the rectum and bladder. Int. J. Radiat. Oncol. Biol. Phys. 20: 1317-1324, 1991. 17 Ten Haken, R.K., Perez-Tamayo C., Tesser R.J., McShan D. L., Fraass, B. A. and Lichter, A.S. Boost treatment of the prostate using shaped, fixed fields. Int. J. Radiat. Oncol. Biol. Phys. 16: 193-200, 1989. 18 Van Herk, M. Physical aspects of a liquid filled ionization chamber with pulsed polarizing voltage. Med. Phys. 18: 692-702, 1991. 19 Van Herk, M. and Meertens, H. A matrix ionization chamber imaging device for on-line patient setup verification during radiotherapy. Radiother. Oncol. 11: 369-378, 1988.

Maximizing setup accuracy using portal images as applied to a conformal boost technique for prostatic cancer.

A design procedure of a patient setup verification protocol based upon frequent digital acquisition of portal images is demonstrated with an applicati...
1MB Sizes 0 Downloads 0 Views