1992. The British Journal of Radiology, 65, 701-709
Image comparison techniques for use with megavoltage imaging systems By *P. M. Evans, DPhil, t J . Q. Gildersleve, FRCR, **E. J. Morton, PhD, *W. Swindell, DSc, *§R. Coles, MSc, *1TM. Ferraro, BSc, tC. Rawlings, DCR, *Z. R. Xiao, MPhil and t J . Dyer, DCR *Joint Department of Physics and tDepartment of Radiotherapy, Royal Marsden Hospital and Institute of Cancer Research, Downs Road, Sutton, Surrey SM2 5PT, UK {Received 9 October 1991 and in revised form 23 January 1991, accepted 12 February 1992) Keywords: Radiotherapy imaging, Patient position verification, Image comparison
Abstract. In this paper we describe software facilities for enabling patient positioning studies using the megavoltage imaging system developed at the Royal Marsden Hospital and Institute of Cancer Research. The study focuses on the use of the system for three purposes: patient position verification (by comparing images taken at treatment simulation with megavoltage images taken at treatment time); reproducibility studies (by analysing a set of megavoltage images); and set-up correction (by adjusting the setup until the megavoltage image obtained at treatment registers with the simulation image). The need is discussed for suitably presented simulator images, a method of determining field boundaries and the possibility of delineating soft-tissue interfaces. Several algorithms of different types, developed specifically for the purpose of intercomparison of planar projection images, are presented. The techniques employed and their usefulness, in both the qualitative and the quantitative sense, are discussed. The results are presented of a phantom and clinical study, to evaluate the rigour and reproducibility of the algorithms. These results indicate that measurements can be made to an accuracy of about 1-2 mm, with a similar value for interobserver reproducibility for the best image comparison techniques available.
Recent studies have revealed errors of up to several centimetres in patient positioning during radiotherapy (Marks etal, 1974; Byhardt etal, 1978; Rabinowitz etal, 1985; Lam etal, 1987; Huizenga etal, 1988; Graham etal, 1991), a magnitude sufficient both to decrease local control and to increase complications. The most commonly employed method of determining the correctness of the patient's positioning is to take checkfilmimages. However, the efficacy of thesefilmsis severely limited. The image quality is poor, because of low contrast and the large amount of scatter. Also, a check film has to be removed from the treatment room and developed before the image can be viewed. With these deficiencies in mind, several groups have developed digital X-ray detection systems for use at megavoltage energies (see, for example, Baily et al, 1980; Meertens etal, 1985; Lam etal, 1986; Leong, 1986; Bova etal, 1987; Leong & Stracher, 1987; Wilenzick etal, 1987; Cheng etal, 1988; Shalev etal, 1988; Reinstein etal, 1988; van Herk & Meertens, 1988; Shalev et al, 1989; Munro et al, 1990; Wong et al, 1990; Antonuk et al, 1991). The intention is that such devices will eventually supersede check films. As the images are digital, they are readily accessible to computerized J Present address: Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, USA. § Present address: Department of Medical Physics and Bioengineering, University Hospital of Wales, Cardiff CF4 4XW, UK. 1{ Present address: Physics Department, Brunei University, Uxbridge, Middlesex UB8 3PH, UK. Vol. 65, No. 776
image processing techniques. Furthermore, images are available within a matter of seconds, rather than minutes as with conventional film. We have developed a digital megavoltage X-ray system of the type described above (Morton & Swindell, 1987; Morton etal, 1991a). A unique feature of our system is the choice of photon detection medium: a scanning array of 128 ZnWO4 scintillation crystals each of which is optically coupled to a silicon photodiode. The detective quantum efficiency of such a system is very high (> 50%), enabling the formation of an image with a relatively small dose (< 1 cGy). However, as the device is a scanning system, the total dose delivered to the patient is approximately 25 cGy per frame (Morton etal, 1991a). The portal scanner is reproducibly placed 1 m from the isocentre of a Philips SL25 linear accelerator, ensuring that the megavoltage images all have the same magnification and that the scatter component is minimal (Swindell et al, 1991). Full details of the design, construction and development of this device are given elsewhere (Morton et al, 1991a). Here we concentrate on work to establish facilities for image comparison using the system. Specifically, a library of software tools for patient image analysis has been developed. Various qualitative and quantitative algorithms have been implemented to enable interimage comparison and hence the determination of set-up and positioning errors. Analyses obtained from such algorithms serve three main purposes: first, comparison of a megavoltage image with a simulation image enables verification of patient positioning relative to the de facto standard; second, comparison of a series of images 701
P. M. Evans et al
taken at various fractions throughout the course of the treatment regime allows one to establish how reproducible is the patient's set-up; and third, if the results obtained from such an image comparison process conform to the mechanical geometry of the treatment machine {e.g. if rotations about the z-axis are calculated about the isocentre, etc.) then such results may be used to correct the patient's positioning during the course of a particular treatment fraction. Clearly, for such a procedure to be practicable, the image comparison process must be both accurate and quick. This work was implemented on a //VAX II computer interfaced to a LEXIDATA 3703 image display system. In the following section the specific requirements for verification and reproducibility studies using a digital imaging system are discussed. In the next section the image comparison algorithms are presented, following which the usefulness of the methods is considered, along with their consistency and ease of use. The conclusions drawn from this study are given in the final section. General considerations Qualitative and quantitative comparison techniques Qualitative techniques of image comparison, although not providing numerical output, still have much use in radiotherapy verification. Analysis of a set of images may yield information on the ways in which a particular treatment procedure is inadequate. For example, it might indicate that there is significant movement in one degree of freedom yet not in another. Quantitative measurement techniques for use in the radiotherapy clinic have to satisfy certain criteria, the most fundamental of which is numerical accuracy. If the minimum resolvable distance of a measurement system is greater in size than the magnitude of the variability of the position of the patient, then clearly this variability cannot be measured. Our system has a spatial resolution of about 3 mm. However, variations smaller than this may be observed in the images owing to the fact that the human eye is able to average information over several pixels. Reported studies suggest that errors of up to 1 cm will be found on a regular basisf(e.g. Lam et al, 1987). The inherently poor contrast in megavoltage energy images makes the use of algorithms to delineate features of interest within the images for automatic analysis difficult. Thus we rely on the user to establish the features of interest. This further optimizes the speed of implementation as required for on-line correction of patient set-up by eliminating the need for complex image-registration software. The scheme used to implement quantitative image comparison is discussed in the Appendix. Treatment simulation images For a radiotherapy portal imaging system to be useful for treatment verification purposes, a reference image has to be available. This will generally be the treatment simulation image. We have described previously how we obtain X-ray image intensifer images from our treat702
ment simulator for incorporation into our imaging system (Morton et al, 1991b). We also digitize simulator films with a CCD camera and use them in the same manner. As our megavoltage scanner is rigidly mounted at a distance of 1 m from the isocentre, the pixel size is constant. Thus, all simulator images may easily be scaled so that they have the same pixel size and, hence, magnification as a megavoltage image. The result is a simulator image that is scaled to the dimensions of a corresponding megavoltage image. Several other groups have discussed the digitization of check film and simulator images (Leong, 1984; Meetens, 1985; Amols & Lowinger, 1987; Grimm et al, 1987). Regions of interest We have a facility for specifying a region of interest (ROI). To draw a region of interest, a cursor is moved around the structure to be delineated, tracing out a wire-frame diagram. Field edge delineation The quantity that must be measured when evaluating patient set-up is the position of the patient's anatomy relative to the field edge, and so, if two images are to be compared, as a first step their relative position must be adjusted until their field edges coincide. This then constitutes the starting point for any measurement of patient position and may be achieved by adjusting the position of two regions of interest, one corresponding to eachfieldoutline, until maximal coincidence is achieved. Any measurements are made from this starting point. This assumes that the absolute positioning of thefieldis unimportant, i.e. that the dose distribution remains constant for small rotations and translations of the collimation system. It is clear from the above discussion that thefieldedge must be found. In the case of the simulator images this is easily achieved usingfield-definingwires which indicate the collimator positioning. A region of interest is drawn manually to trace out the field edge of the simulator and then superimposed on the megavoltage image. The field edge of the megavoltage image is determined (as described below) and also superimposed on the megavoltage image as a region of interest. The operator then adjusts the positioning of the simulator field edge until the two field edges coincide maximally. The adjustment required to achieve this is never more than two pixels. Since the simulator images are obtained by placing a film in front of a camera or from an image intensifier, neither of whose position need be fixed, the above procedure is needed to ensure the simulator field edge in the image registers exactly with the expected megavoltage image field edge. The rigid fixation of our imaging system then enables absolute measurement of field edge positioning as well as patient positioning. The problem of finding the field edge on a megavoltage image has been discussed in the literature where a dual-exposure technique has been presented (Meertens et al, 1990a,b). We have tried to establish the field edge without the need for dual exposure. The field edge is The British Journal of Radiology, August 1992
Image comparison techniques
variations in machine output (particularly fluctuations across the whole field) mean that a given anatomical feature in one image may have a different intensity value compared with the corresponding feature in another image, a fact that compounds the problem of determining soft-tissue interfaces. We have approached this problem by applying histogram equalization to each image to be measured. The assumption is then made that a given anatomical feature has a constant, renormalized intensity in each image. This is only valid if the histogram for each image would be the same were the intensity values identical, a hypothesis that requires that the patient positioning in each image be very similar, if not identical. This method is only valid for comparing megavoltage images with other megavoltage images, as the relative strengths of the Compton scattering and photo-electric interaction processes are different at the diagnostic energies used to produce simulation images, resulting in different image quality. Image analysis algorithms
Figure 1. The field edge, defined at the 50% isodose level, determined for a "dogleg" field. Also shown is the use of the ruler technique to establish the distance between the vertebral column and the field edge.
defined at the 50% isodose level. We have used a simple method based on this to estimate the field edge. The brightest region of the image is analysed and the mean intensity there deduced and taken as the 100% level. The image is then scanned line-by-line and the 50% intensity contour established. This contour is then turned into a region of interest for further manipulation. This method relies on the assumption that the distribution of attenuation across the field is uniform. A small amount of non-uniformity will still allow accurate results, as the 50% level coincides with the steepest part of the penumbra cast by the field edge. Hence, a small error in the 100% level still yields the same position for the contour of the 50% level. In order to determine the limitations of this algorithm we have taken a set of phantom images corresponding to an extreme situation—a set of square fields with a 12 cm thickness of Perspex at one end and open field at the other. The error on the measured field edge was no more than 3 mm in the worst case. In practice, however, the calculated field size is in agreement with the collimators' prescribed settings to 1.5 mm or better. In Fig. 1 we present afieldedge determination using this method for a "dogleg" field. This technique has been applied to some wedged fields and has yielded results of a similar accuracy to the phantom study described above. Soft-tissue interfaces The apparent positioning of soft-tissue/air and softtissue/bone interfaces depends critically on the greyscale window of the displayed image. Furthermore, Wol. 65, No. 776
Qualitative methods Side-by-side. This method involves the display of two megavoltage scans, or a megavoltage image and a digitized simulator image, next to one another. Gross differences, such as a large discrepancy in rotation or the absence or misplacing of shielding blocks, would be recognizable immediately. One may also superimpose a chosen region of interest over each image. Deviations between the region of interest outline and the images would indicate the nature of any set-up errors relative to the reference image. This method is included in our system as it is traditionally employed to compare verification and simulator film images manually. Movies. The human eye is very sensitive to motion. This property of the eye is exploited in the movie procedure. A set of images is displayed, in a temporal sequence, at the centre of the image display system's screen (Reinstein etal, 1988). Differences in anatomical structures and field definition edges between two images are vividly perceived as rapid movement. Fading between images. In this method two images are displayed superimposed over one another in the form of a composite image whose neighbouring pixel values alternate between those of the first image and those of the second image in a "checker board" fashion. Furthermore, the pixels determined from the second image are given a DC offset so they are written in a different part of the LEXIDATA's grey-scale look-up table to those from the first image, to enable the two components to be windowed independently. The composite image is then windowed in the following manner: as the mean window level for pixels taken from the first image increases, the level for pixels from the second image decreases and vice versa. Thus, as the windowing of the composite image is adjusted, the display visually changes from one input image to the other. In between these two extremes the two are displayed superimposed and with similar brightness. A megavoltage image displayed in this manner superimposed on the corre703
P. M. Evans et al
Figure 3. The graphical user interface of feature transposition.
is achieved. The shift of the ROI's centre and its rotation (about the isocentre) are displayed (Fig. 4). This routine enables a quick and simple determination sponding simulator image for a pelvic field is shown in of the position of the patient relative to the reference. No magnification parameter is incorporated in this Fig. 2. algorithm at present. However, as pointed out above, all images should have the same magnification. Quantitative methods Movie measurement. The philosophy of the movie Distance and angle measurement. Conventionally, when two images are compared a ruler and a protractor procedure described above has been applied to the are used to measure distances and angles mechanically. problem of quantitative measurement. The image to be Such facilities are available on our system. To measure a measured and the reference image are displayed in a distance the operator places a cursor at two points of movie consisting of two images that are displayed alterinterest. The distance between these points is then calcu- nately at a rate of typically 6 frames per second. During lated. Such a tool may be used to establish the distances the course of the movie, the operator may adjust the between critical structures and the field edge, for position of the image to be measured by rotation and example (Fig. 1). A similar facility exists for measuring translation. This is continued iteratively until the angles. Two lines are specified and the angle between features of interest, such as bony structures, are seen to become stationary as the display oscillates from one them is evaluated. Feature transposition. Two images are displayed side- image to the other. Translations are achieved, in real by-side. A point is marked on one image along with the corresponding point on the other, using a cursor. This procedure is repeated until several points of interest have been indicated. The centre shift (xt, >>,), rotation (0t) and magnification (m^ my) of one image relative to the other are then calculated using the formulae presented in Equations 1-3 (Appendix). This method should give the same results as all of the methods presented below and should indicate unit magnification for all images used on our system. A method similar to this has been implemented by Meertens et al (1990a), as discussed above. The user-interface of this facility is shown in Fig. 3. In practice, the two images are displaced in the y direction to prevent any biasing of results. Region of interest transposition. In this method a region of interest (obtained from a reference image) is superimposed over the image being analysed. The operator then uses a trackerball interfaced to the LEXIDATA display system to move and rotate the Figure 4. The graphical user interface of region of interest ROI until the best registration with the analysed image transposition. Figure 2. A fade image comprising a simulator image and the corresponding treatment time, mega voltage scan.
704
The British Journal of Radiology, August 1992
Image comparison techniques
Figure 5. A difference image display. The top two images are under evaluation; the bottom right-hand image is a movie (see text), and the bottom left-hand image is the difference image. The blue region is to the left of the mediastinum (i.e. on the right-hand side of the difference image) and the red region is to the right of the mediastinum.
Figure 6. The corresponding image to Fig. 5 with maximal alignment of anatomy. The coloured regions at the edge of the difference image indicate that the collimators now appear in a different position and are caused by penumbra.
threshold settings, because of the large penumbra region associated with the edge of such features (see Fig. 6). time, by adjusting the position at which the image is displayed on the LEXIDATA's screen. All rotations have to be carried out using digital resampling and hence are slightly slower (typically 3 s). Difference images. The reference image and the image to be measured are normalized, using histogram equalization, as above. A pixel-by-pixel subtraction is then carried out of one image from the other. If the intensity of thefirstimage exceeds that of the second one by more than a chosen threshold value, typically 3% of the mean image intensity, then the difference image is coloured red. Conversely, if the intensity of the second image is greater by more than the threshold, then it is coloured blue. To enable the observer to relate the positioning of these coloured regions to the topology of the original images, if the difference is below the threshold value, then the intensity of the first image is ascribed to the difference image. The two input images are also displayed. The operator then adjusts the relative positioning of the images until the amount of colour present in the anatomical detail within the difference image is minimized. This method cannot be used to compare megavoltage images with simulator images, as the dominant photon interaction process is different for the two image types, resulting in significantly different relative pixel intensities. Figure 5 shows a difference image before image position correction and Fig. 6 shows the same image once the anatomical detail has been aligned. The source size of a radiotherapy linear accelerator is typically 2 mm at the 4% radius and the primary collimation system is of the order of 40 cm from the source. Our detector is located 2 m from the source and thus the penumbra cast will be of the order of 8 mm. The difference image method will automatically highlight block positioning and field size errors, even for the highest Vol. 65, No. 776
Evaluation
We grouped the measurement algorithms into two categories, depending on whether the information provided is of a qualitative or a quantitative nature. Of the five quantitative methods presented, four are for the intercomparison of two images, whereas the distance, angle technique provides information on a single image. The intercomparison techniques may readily be verified by using them to compare images whose differences are known a priori. To this end a humanoid phantom was imaged at a set of known rotations, and transverse and longitudinal positions, which were obtained by adjusting the position of the treatment couch. The methods of feature transposition, ROI transposition, movie measurement and difference images were applied by four observers to measure the differences between a reference image and five measurement images. A typical set of results for such a measurement is shown in Table I. The symbol \8\ in Table I denotes the average of the moduli of the difference measurements in x, y and 6 for each technique and thus gives an estimate of the error in each case. This suggests that the average error in angle is far less than 1 ° and that the average error in JC or y is no more than 1 mm. Although the above results suggest that all four techniques fulfil the requirements of accuracy and reliability, the use of humanoid phantom images, in the manner described above, is somewhat different from the real situation encountered in the clinic, in that the phantom is always in the same position relative to the treatment couch. Hence, a perfect match should always be achievable simply by rotating and translating the treatment couch or treatment head (assuming the gantry to be at 0°). However, the position of a real patient on the 705
P. M. Evans et at* Table I. A typical observer's results for the measured deviations from the reference image obtained for a set of phantom images, x, and yt are tabulated in millimetres and 0t is in degrees. "True" denotes the correct difference values. \d\ is the average of the moduli of the differences between the values in each column and the correct value; it thus indicates the relative errors associated with each method Movie
Image True
1 2 3 4 5
Difference
ROI trans.
Feature trans.
xt
yt
0.
xt
yt
ot
xt
yx
0,
xt
yt
et
xt
J>t
0,
0 0 13 13 13
0 9 9 5 6
1 -3 -3 -2 4
0 -I 12 12 12
0 JO
1 -3 -3 -2 4
0 0 12 11 12
0 9 9 5 5
1 -3 -3 -3 3
1 2 13 16 11
-l 9 11 6 5
1 -2 -2 -2 5
1 0 11 13 12
-1 11 10 6 5
0 -1 -3 -2 4
1*1
0.8
n6 4 1.2
0.0
0.8
0.2
treatment couch will be subject to variations due to rotations out of plane and to anatomical movements, factors that may make a good match between different treatment-time images improbable, if not impossible. Clearly, for a set of real, treatment-time patient images, the variations in position between different images is unknown. However, reproducibility between observers may be tested and is likely to be a good indication of accuracy. To this end, the four techniques have been applied to the analysis of a set of anterior images obtained from a patient being treated in the pelvic region. Four observers made measurements of the differences between a reference image and three other images of this patient. One of the images used is shown in Fig. 4. The results obtained are presented in Table II. For each pair of images, the last row of Table II lists values for the mean deviation and thus is an indicator of
0.4
1.6
0.6
1.0
0.8
0.6
1.2
reproducibility. Clearly, the movie movement technique displays the best reproducibility. Of the four intercomparison methods, the movie measurement technique was found to produce the most reproducible results, followed by the difference image, then ROI transposition and feature transposition. The good reproducibility of the movie measurement procedure is brought about by its use of all of the information within the image. Furthermore, it does not require user input to create wire-frame diagrams or to choose specific points of interest, thus removing potential sources of error. For the difference image, because of systematic noise (such as vertical streaks caused by drifts in detector gain) and a non-standard mode of image presentation, it was found necessary to vary the amount of shift quite significantly in order for the observer to establish the component of colour informa-
Table II. Results obtained from the intercomparison of a sequence of clinical images of a patient being treated in the pelvic region (see Fig. 4). Values for the mean deviation obtained from each measurement technique (a) are also presented, x, and yt are in millimetres and 0, is in degrees. Image
1
Observer
1 2 3 4
1 2 3 4
a 3
a
706
1 2 3 4
ROI trans. xt
3>t
0,
xt
yt
0,
0 -2 1 2
-1 -2 0 -4
-1 0 -2 -2
0 0 -1 _2
-2 0 -3 0
-1 -2 -1 -2
yt
et
xt
yt
-2 -2 -2 -2
-3 -3 -3 -3
-2 -2 -2 -2
-2 -3 -2 -3
-3 -2 -3 -3
2 2 3 3
0.0 _3 -3 -2 -3
0.6
0.5
3 3 3 3
5 3 5 3
0.0
1.2
0.5
0.0
0.6
1 1 1 1
3 2 3 3
0.0
0.5
0.6
3 3 3 5
5 5 5 5
1.0
0.0
-1 0 0 0 0.5
Feature trans.
Difference
xt
0.0
a 2
Movie
-2 -3 -2 -3
-1 0 -2 0 1.0
1.7
1 2 0 1
2 4 3 1
0.8
1.3
1.5
4 3 4 3
3 2 2 4
0.6
1.0
-2 -2 -1 2 0.8
1.7 -2 -2 -2 -5
1.5
0.6
1.0
1.0
2 2 1 3
4 3 2 1
0.8
1.3
2.9
1.3
3 3 5 3
4 5 5 4
0 0 3 0
1.0
0.6
1.5
-1 0 -1 0 0.6
1 4 -2 -2
2 -1 1 1
The British Journal of Radiology, August 1992
Image comparison techniques
tion resulting from systematic noise and to determine which of the colour information is due to shifts and which is due to rotations. Thus, the difference image tool was found to be slow to use for evaluating anatomical shifts. To this end, a movie of the images has been added (Figs 5, 6). Owing to the penumbra effect, this method is, however, extremely sensitive to small variations in field position. The field edge detection algorithm was evaluated using another set of phantom images, where the field size was set to a series of known values. 10 measurements were taken for each field size. The results indicate an error of no greater than one pixel (1.5 mm) and a standard deviation of typically no more than 0.5 mm. It is not possible to present a numerical evaluation of the qualitative tools, so we discuss them subjectively. The movie procedure was found to be by far the most useful. It provides good visualization of shifts and rotations of both anatomical structures and the treatment field. Furthermore, it is easy to spotlight features for measurement using a movie sequence. The side-byside image comparison method was found to be less valuable than the two other qualitative methods, unless used in conjunction with an ROI. In this case it provides a means of detecting image offsets very quickly, unlike the movie technique, which requires time to set up the sequence of images. Rotations are extremely difficult to detect with this method. The process of fading between images has been found to be useful when comparing megavoltage and simulator images, particularly as the simulator images contain information beyond the field edge, whereas the megavoltage images do not. Summary and conclusions
In this paper we have presented an evaluation of the RMH/ICR megavoltage imaging system for patient position reproducibility studies. Various key considerations pertinent to the use of the system to make measurements have been presented: the need to have access to simulator images; the problem of finding and matching field edges; the delineation of soft-tissue interfaces; the efficacy of using regions of interest; and the desirable properties of a quantitative measurement procedure. Several algorithms for image measurement were presented. Four provide image intercomparison information: feature transposition, ROI transposition, movie measurement, and difference images. Three are qualitative: movie, side-by-side, and fading between images. A ruler, angle measurement tool was also discussed. The evaluation of the intercomparison algorithms shows that movie measurement, using the whole image, was the best method, followed by difference images, followed by ROI transposition, typically using the edge of a structure, followed by feature transposition, using a set of points. Thus, we conclude that the usefulness of these algorithms is determined by how much of the information content of the images is used. The difference image technique is useful for determining shifts in l. 65, No. 776
megavoltage images, but requires much time to implement (typically several minutes compared with 1 min or less for the other techniques). It also requires significant user familiarity because of its non-standard presentation of the image data. In conclusion, our megavoltage imaging system is intended for patient position verification studies. A short image acquisition time (
Image comparison techniques for use with megavoltage imaging systems By *P. M. Evans, DPhil, t J . Q. Gildersleve, FRCR, **E. J. Morton, PhD, *W. Swindell, DSc, *§R. Coles, MSc, *1TM. Ferraro, BSc, tC. Rawlings, DCR, *Z. R. Xiao, MPhil and t J . Dyer, DCR *Joint Department of Physics and tDepartment of Radiotherapy, Royal Marsden Hospital and Institute of Cancer Research, Downs Road, Sutton, Surrey SM2 5PT, UK {Received 9 October 1991 and in revised form 23 January 1991, accepted 12 February 1992) Keywords: Radiotherapy imaging, Patient position verification, Image comparison
Abstract. In this paper we describe software facilities for enabling patient positioning studies using the megavoltage imaging system developed at the Royal Marsden Hospital and Institute of Cancer Research. The study focuses on the use of the system for three purposes: patient position verification (by comparing images taken at treatment simulation with megavoltage images taken at treatment time); reproducibility studies (by analysing a set of megavoltage images); and set-up correction (by adjusting the setup until the megavoltage image obtained at treatment registers with the simulation image). The need is discussed for suitably presented simulator images, a method of determining field boundaries and the possibility of delineating soft-tissue interfaces. Several algorithms of different types, developed specifically for the purpose of intercomparison of planar projection images, are presented. The techniques employed and their usefulness, in both the qualitative and the quantitative sense, are discussed. The results are presented of a phantom and clinical study, to evaluate the rigour and reproducibility of the algorithms. These results indicate that measurements can be made to an accuracy of about 1-2 mm, with a similar value for interobserver reproducibility for the best image comparison techniques available.
Recent studies have revealed errors of up to several centimetres in patient positioning during radiotherapy (Marks etal, 1974; Byhardt etal, 1978; Rabinowitz etal, 1985; Lam etal, 1987; Huizenga etal, 1988; Graham etal, 1991), a magnitude sufficient both to decrease local control and to increase complications. The most commonly employed method of determining the correctness of the patient's positioning is to take checkfilmimages. However, the efficacy of thesefilmsis severely limited. The image quality is poor, because of low contrast and the large amount of scatter. Also, a check film has to be removed from the treatment room and developed before the image can be viewed. With these deficiencies in mind, several groups have developed digital X-ray detection systems for use at megavoltage energies (see, for example, Baily et al, 1980; Meertens etal, 1985; Lam etal, 1986; Leong, 1986; Bova etal, 1987; Leong & Stracher, 1987; Wilenzick etal, 1987; Cheng etal, 1988; Shalev etal, 1988; Reinstein etal, 1988; van Herk & Meertens, 1988; Shalev et al, 1989; Munro et al, 1990; Wong et al, 1990; Antonuk et al, 1991). The intention is that such devices will eventually supersede check films. As the images are digital, they are readily accessible to computerized J Present address: Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, USA. § Present address: Department of Medical Physics and Bioengineering, University Hospital of Wales, Cardiff CF4 4XW, UK. 1{ Present address: Physics Department, Brunei University, Uxbridge, Middlesex UB8 3PH, UK. Vol. 65, No. 776
image processing techniques. Furthermore, images are available within a matter of seconds, rather than minutes as with conventional film. We have developed a digital megavoltage X-ray system of the type described above (Morton & Swindell, 1987; Morton etal, 1991a). A unique feature of our system is the choice of photon detection medium: a scanning array of 128 ZnWO4 scintillation crystals each of which is optically coupled to a silicon photodiode. The detective quantum efficiency of such a system is very high (> 50%), enabling the formation of an image with a relatively small dose (< 1 cGy). However, as the device is a scanning system, the total dose delivered to the patient is approximately 25 cGy per frame (Morton etal, 1991a). The portal scanner is reproducibly placed 1 m from the isocentre of a Philips SL25 linear accelerator, ensuring that the megavoltage images all have the same magnification and that the scatter component is minimal (Swindell et al, 1991). Full details of the design, construction and development of this device are given elsewhere (Morton et al, 1991a). Here we concentrate on work to establish facilities for image comparison using the system. Specifically, a library of software tools for patient image analysis has been developed. Various qualitative and quantitative algorithms have been implemented to enable interimage comparison and hence the determination of set-up and positioning errors. Analyses obtained from such algorithms serve three main purposes: first, comparison of a megavoltage image with a simulation image enables verification of patient positioning relative to the de facto standard; second, comparison of a series of images 701
P. M. Evans et al
taken at various fractions throughout the course of the treatment regime allows one to establish how reproducible is the patient's set-up; and third, if the results obtained from such an image comparison process conform to the mechanical geometry of the treatment machine {e.g. if rotations about the z-axis are calculated about the isocentre, etc.) then such results may be used to correct the patient's positioning during the course of a particular treatment fraction. Clearly, for such a procedure to be practicable, the image comparison process must be both accurate and quick. This work was implemented on a //VAX II computer interfaced to a LEXIDATA 3703 image display system. In the following section the specific requirements for verification and reproducibility studies using a digital imaging system are discussed. In the next section the image comparison algorithms are presented, following which the usefulness of the methods is considered, along with their consistency and ease of use. The conclusions drawn from this study are given in the final section. General considerations Qualitative and quantitative comparison techniques Qualitative techniques of image comparison, although not providing numerical output, still have much use in radiotherapy verification. Analysis of a set of images may yield information on the ways in which a particular treatment procedure is inadequate. For example, it might indicate that there is significant movement in one degree of freedom yet not in another. Quantitative measurement techniques for use in the radiotherapy clinic have to satisfy certain criteria, the most fundamental of which is numerical accuracy. If the minimum resolvable distance of a measurement system is greater in size than the magnitude of the variability of the position of the patient, then clearly this variability cannot be measured. Our system has a spatial resolution of about 3 mm. However, variations smaller than this may be observed in the images owing to the fact that the human eye is able to average information over several pixels. Reported studies suggest that errors of up to 1 cm will be found on a regular basisf(e.g. Lam et al, 1987). The inherently poor contrast in megavoltage energy images makes the use of algorithms to delineate features of interest within the images for automatic analysis difficult. Thus we rely on the user to establish the features of interest. This further optimizes the speed of implementation as required for on-line correction of patient set-up by eliminating the need for complex image-registration software. The scheme used to implement quantitative image comparison is discussed in the Appendix. Treatment simulation images For a radiotherapy portal imaging system to be useful for treatment verification purposes, a reference image has to be available. This will generally be the treatment simulation image. We have described previously how we obtain X-ray image intensifer images from our treat702
ment simulator for incorporation into our imaging system (Morton et al, 1991b). We also digitize simulator films with a CCD camera and use them in the same manner. As our megavoltage scanner is rigidly mounted at a distance of 1 m from the isocentre, the pixel size is constant. Thus, all simulator images may easily be scaled so that they have the same pixel size and, hence, magnification as a megavoltage image. The result is a simulator image that is scaled to the dimensions of a corresponding megavoltage image. Several other groups have discussed the digitization of check film and simulator images (Leong, 1984; Meetens, 1985; Amols & Lowinger, 1987; Grimm et al, 1987). Regions of interest We have a facility for specifying a region of interest (ROI). To draw a region of interest, a cursor is moved around the structure to be delineated, tracing out a wire-frame diagram. Field edge delineation The quantity that must be measured when evaluating patient set-up is the position of the patient's anatomy relative to the field edge, and so, if two images are to be compared, as a first step their relative position must be adjusted until their field edges coincide. This then constitutes the starting point for any measurement of patient position and may be achieved by adjusting the position of two regions of interest, one corresponding to eachfieldoutline, until maximal coincidence is achieved. Any measurements are made from this starting point. This assumes that the absolute positioning of thefieldis unimportant, i.e. that the dose distribution remains constant for small rotations and translations of the collimation system. It is clear from the above discussion that thefieldedge must be found. In the case of the simulator images this is easily achieved usingfield-definingwires which indicate the collimator positioning. A region of interest is drawn manually to trace out the field edge of the simulator and then superimposed on the megavoltage image. The field edge of the megavoltage image is determined (as described below) and also superimposed on the megavoltage image as a region of interest. The operator then adjusts the positioning of the simulator field edge until the two field edges coincide maximally. The adjustment required to achieve this is never more than two pixels. Since the simulator images are obtained by placing a film in front of a camera or from an image intensifier, neither of whose position need be fixed, the above procedure is needed to ensure the simulator field edge in the image registers exactly with the expected megavoltage image field edge. The rigid fixation of our imaging system then enables absolute measurement of field edge positioning as well as patient positioning. The problem of finding the field edge on a megavoltage image has been discussed in the literature where a dual-exposure technique has been presented (Meertens et al, 1990a,b). We have tried to establish the field edge without the need for dual exposure. The field edge is The British Journal of Radiology, August 1992
Image comparison techniques
variations in machine output (particularly fluctuations across the whole field) mean that a given anatomical feature in one image may have a different intensity value compared with the corresponding feature in another image, a fact that compounds the problem of determining soft-tissue interfaces. We have approached this problem by applying histogram equalization to each image to be measured. The assumption is then made that a given anatomical feature has a constant, renormalized intensity in each image. This is only valid if the histogram for each image would be the same were the intensity values identical, a hypothesis that requires that the patient positioning in each image be very similar, if not identical. This method is only valid for comparing megavoltage images with other megavoltage images, as the relative strengths of the Compton scattering and photo-electric interaction processes are different at the diagnostic energies used to produce simulation images, resulting in different image quality. Image analysis algorithms
Figure 1. The field edge, defined at the 50% isodose level, determined for a "dogleg" field. Also shown is the use of the ruler technique to establish the distance between the vertebral column and the field edge.
defined at the 50% isodose level. We have used a simple method based on this to estimate the field edge. The brightest region of the image is analysed and the mean intensity there deduced and taken as the 100% level. The image is then scanned line-by-line and the 50% intensity contour established. This contour is then turned into a region of interest for further manipulation. This method relies on the assumption that the distribution of attenuation across the field is uniform. A small amount of non-uniformity will still allow accurate results, as the 50% level coincides with the steepest part of the penumbra cast by the field edge. Hence, a small error in the 100% level still yields the same position for the contour of the 50% level. In order to determine the limitations of this algorithm we have taken a set of phantom images corresponding to an extreme situation—a set of square fields with a 12 cm thickness of Perspex at one end and open field at the other. The error on the measured field edge was no more than 3 mm in the worst case. In practice, however, the calculated field size is in agreement with the collimators' prescribed settings to 1.5 mm or better. In Fig. 1 we present afieldedge determination using this method for a "dogleg" field. This technique has been applied to some wedged fields and has yielded results of a similar accuracy to the phantom study described above. Soft-tissue interfaces The apparent positioning of soft-tissue/air and softtissue/bone interfaces depends critically on the greyscale window of the displayed image. Furthermore, Wol. 65, No. 776
Qualitative methods Side-by-side. This method involves the display of two megavoltage scans, or a megavoltage image and a digitized simulator image, next to one another. Gross differences, such as a large discrepancy in rotation or the absence or misplacing of shielding blocks, would be recognizable immediately. One may also superimpose a chosen region of interest over each image. Deviations between the region of interest outline and the images would indicate the nature of any set-up errors relative to the reference image. This method is included in our system as it is traditionally employed to compare verification and simulator film images manually. Movies. The human eye is very sensitive to motion. This property of the eye is exploited in the movie procedure. A set of images is displayed, in a temporal sequence, at the centre of the image display system's screen (Reinstein etal, 1988). Differences in anatomical structures and field definition edges between two images are vividly perceived as rapid movement. Fading between images. In this method two images are displayed superimposed over one another in the form of a composite image whose neighbouring pixel values alternate between those of the first image and those of the second image in a "checker board" fashion. Furthermore, the pixels determined from the second image are given a DC offset so they are written in a different part of the LEXIDATA's grey-scale look-up table to those from the first image, to enable the two components to be windowed independently. The composite image is then windowed in the following manner: as the mean window level for pixels taken from the first image increases, the level for pixels from the second image decreases and vice versa. Thus, as the windowing of the composite image is adjusted, the display visually changes from one input image to the other. In between these two extremes the two are displayed superimposed and with similar brightness. A megavoltage image displayed in this manner superimposed on the corre703
P. M. Evans et al
Figure 3. The graphical user interface of feature transposition.
is achieved. The shift of the ROI's centre and its rotation (about the isocentre) are displayed (Fig. 4). This routine enables a quick and simple determination sponding simulator image for a pelvic field is shown in of the position of the patient relative to the reference. No magnification parameter is incorporated in this Fig. 2. algorithm at present. However, as pointed out above, all images should have the same magnification. Quantitative methods Movie measurement. The philosophy of the movie Distance and angle measurement. Conventionally, when two images are compared a ruler and a protractor procedure described above has been applied to the are used to measure distances and angles mechanically. problem of quantitative measurement. The image to be Such facilities are available on our system. To measure a measured and the reference image are displayed in a distance the operator places a cursor at two points of movie consisting of two images that are displayed alterinterest. The distance between these points is then calcu- nately at a rate of typically 6 frames per second. During lated. Such a tool may be used to establish the distances the course of the movie, the operator may adjust the between critical structures and the field edge, for position of the image to be measured by rotation and example (Fig. 1). A similar facility exists for measuring translation. This is continued iteratively until the angles. Two lines are specified and the angle between features of interest, such as bony structures, are seen to become stationary as the display oscillates from one them is evaluated. Feature transposition. Two images are displayed side- image to the other. Translations are achieved, in real by-side. A point is marked on one image along with the corresponding point on the other, using a cursor. This procedure is repeated until several points of interest have been indicated. The centre shift (xt, >>,), rotation (0t) and magnification (m^ my) of one image relative to the other are then calculated using the formulae presented in Equations 1-3 (Appendix). This method should give the same results as all of the methods presented below and should indicate unit magnification for all images used on our system. A method similar to this has been implemented by Meertens et al (1990a), as discussed above. The user-interface of this facility is shown in Fig. 3. In practice, the two images are displaced in the y direction to prevent any biasing of results. Region of interest transposition. In this method a region of interest (obtained from a reference image) is superimposed over the image being analysed. The operator then uses a trackerball interfaced to the LEXIDATA display system to move and rotate the Figure 4. The graphical user interface of region of interest ROI until the best registration with the analysed image transposition. Figure 2. A fade image comprising a simulator image and the corresponding treatment time, mega voltage scan.
704
The British Journal of Radiology, August 1992
Image comparison techniques
Figure 5. A difference image display. The top two images are under evaluation; the bottom right-hand image is a movie (see text), and the bottom left-hand image is the difference image. The blue region is to the left of the mediastinum (i.e. on the right-hand side of the difference image) and the red region is to the right of the mediastinum.
Figure 6. The corresponding image to Fig. 5 with maximal alignment of anatomy. The coloured regions at the edge of the difference image indicate that the collimators now appear in a different position and are caused by penumbra.
threshold settings, because of the large penumbra region associated with the edge of such features (see Fig. 6). time, by adjusting the position at which the image is displayed on the LEXIDATA's screen. All rotations have to be carried out using digital resampling and hence are slightly slower (typically 3 s). Difference images. The reference image and the image to be measured are normalized, using histogram equalization, as above. A pixel-by-pixel subtraction is then carried out of one image from the other. If the intensity of thefirstimage exceeds that of the second one by more than a chosen threshold value, typically 3% of the mean image intensity, then the difference image is coloured red. Conversely, if the intensity of the second image is greater by more than the threshold, then it is coloured blue. To enable the observer to relate the positioning of these coloured regions to the topology of the original images, if the difference is below the threshold value, then the intensity of the first image is ascribed to the difference image. The two input images are also displayed. The operator then adjusts the relative positioning of the images until the amount of colour present in the anatomical detail within the difference image is minimized. This method cannot be used to compare megavoltage images with simulator images, as the dominant photon interaction process is different for the two image types, resulting in significantly different relative pixel intensities. Figure 5 shows a difference image before image position correction and Fig. 6 shows the same image once the anatomical detail has been aligned. The source size of a radiotherapy linear accelerator is typically 2 mm at the 4% radius and the primary collimation system is of the order of 40 cm from the source. Our detector is located 2 m from the source and thus the penumbra cast will be of the order of 8 mm. The difference image method will automatically highlight block positioning and field size errors, even for the highest Vol. 65, No. 776
Evaluation
We grouped the measurement algorithms into two categories, depending on whether the information provided is of a qualitative or a quantitative nature. Of the five quantitative methods presented, four are for the intercomparison of two images, whereas the distance, angle technique provides information on a single image. The intercomparison techniques may readily be verified by using them to compare images whose differences are known a priori. To this end a humanoid phantom was imaged at a set of known rotations, and transverse and longitudinal positions, which were obtained by adjusting the position of the treatment couch. The methods of feature transposition, ROI transposition, movie measurement and difference images were applied by four observers to measure the differences between a reference image and five measurement images. A typical set of results for such a measurement is shown in Table I. The symbol \8\ in Table I denotes the average of the moduli of the difference measurements in x, y and 6 for each technique and thus gives an estimate of the error in each case. This suggests that the average error in angle is far less than 1 ° and that the average error in JC or y is no more than 1 mm. Although the above results suggest that all four techniques fulfil the requirements of accuracy and reliability, the use of humanoid phantom images, in the manner described above, is somewhat different from the real situation encountered in the clinic, in that the phantom is always in the same position relative to the treatment couch. Hence, a perfect match should always be achievable simply by rotating and translating the treatment couch or treatment head (assuming the gantry to be at 0°). However, the position of a real patient on the 705
P. M. Evans et at* Table I. A typical observer's results for the measured deviations from the reference image obtained for a set of phantom images, x, and yt are tabulated in millimetres and 0t is in degrees. "True" denotes the correct difference values. \d\ is the average of the moduli of the differences between the values in each column and the correct value; it thus indicates the relative errors associated with each method Movie
Image True
1 2 3 4 5
Difference
ROI trans.
Feature trans.
xt
yt
0.
xt
yt
ot
xt
yx
0,
xt
yt
et
xt
J>t
0,
0 0 13 13 13
0 9 9 5 6
1 -3 -3 -2 4
0 -I 12 12 12
0 JO
1 -3 -3 -2 4
0 0 12 11 12
0 9 9 5 5
1 -3 -3 -3 3
1 2 13 16 11
-l 9 11 6 5
1 -2 -2 -2 5
1 0 11 13 12
-1 11 10 6 5
0 -1 -3 -2 4
1*1
0.8
n6 4 1.2
0.0
0.8
0.2
treatment couch will be subject to variations due to rotations out of plane and to anatomical movements, factors that may make a good match between different treatment-time images improbable, if not impossible. Clearly, for a set of real, treatment-time patient images, the variations in position between different images is unknown. However, reproducibility between observers may be tested and is likely to be a good indication of accuracy. To this end, the four techniques have been applied to the analysis of a set of anterior images obtained from a patient being treated in the pelvic region. Four observers made measurements of the differences between a reference image and three other images of this patient. One of the images used is shown in Fig. 4. The results obtained are presented in Table II. For each pair of images, the last row of Table II lists values for the mean deviation and thus is an indicator of
0.4
1.6
0.6
1.0
0.8
0.6
1.2
reproducibility. Clearly, the movie movement technique displays the best reproducibility. Of the four intercomparison methods, the movie measurement technique was found to produce the most reproducible results, followed by the difference image, then ROI transposition and feature transposition. The good reproducibility of the movie measurement procedure is brought about by its use of all of the information within the image. Furthermore, it does not require user input to create wire-frame diagrams or to choose specific points of interest, thus removing potential sources of error. For the difference image, because of systematic noise (such as vertical streaks caused by drifts in detector gain) and a non-standard mode of image presentation, it was found necessary to vary the amount of shift quite significantly in order for the observer to establish the component of colour informa-
Table II. Results obtained from the intercomparison of a sequence of clinical images of a patient being treated in the pelvic region (see Fig. 4). Values for the mean deviation obtained from each measurement technique (a) are also presented, x, and yt are in millimetres and 0, is in degrees. Image
1
Observer
1 2 3 4
1 2 3 4
a 3
a
706
1 2 3 4
ROI trans. xt
3>t
0,
xt
yt
0,
0 -2 1 2
-1 -2 0 -4
-1 0 -2 -2
0 0 -1 _2
-2 0 -3 0
-1 -2 -1 -2
yt
et
xt
yt
-2 -2 -2 -2
-3 -3 -3 -3
-2 -2 -2 -2
-2 -3 -2 -3
-3 -2 -3 -3
2 2 3 3
0.0 _3 -3 -2 -3
0.6
0.5
3 3 3 3
5 3 5 3
0.0
1.2
0.5
0.0
0.6
1 1 1 1
3 2 3 3
0.0
0.5
0.6
3 3 3 5
5 5 5 5
1.0
0.0
-1 0 0 0 0.5
Feature trans.
Difference
xt
0.0
a 2
Movie
-2 -3 -2 -3
-1 0 -2 0 1.0
1.7
1 2 0 1
2 4 3 1
0.8
1.3
1.5
4 3 4 3
3 2 2 4
0.6
1.0
-2 -2 -1 2 0.8
1.7 -2 -2 -2 -5
1.5
0.6
1.0
1.0
2 2 1 3
4 3 2 1
0.8
1.3
2.9
1.3
3 3 5 3
4 5 5 4
0 0 3 0
1.0
0.6
1.5
-1 0 -1 0 0.6
1 4 -2 -2
2 -1 1 1
The British Journal of Radiology, August 1992
Image comparison techniques
tion resulting from systematic noise and to determine which of the colour information is due to shifts and which is due to rotations. Thus, the difference image tool was found to be slow to use for evaluating anatomical shifts. To this end, a movie of the images has been added (Figs 5, 6). Owing to the penumbra effect, this method is, however, extremely sensitive to small variations in field position. The field edge detection algorithm was evaluated using another set of phantom images, where the field size was set to a series of known values. 10 measurements were taken for each field size. The results indicate an error of no greater than one pixel (1.5 mm) and a standard deviation of typically no more than 0.5 mm. It is not possible to present a numerical evaluation of the qualitative tools, so we discuss them subjectively. The movie procedure was found to be by far the most useful. It provides good visualization of shifts and rotations of both anatomical structures and the treatment field. Furthermore, it is easy to spotlight features for measurement using a movie sequence. The side-byside image comparison method was found to be less valuable than the two other qualitative methods, unless used in conjunction with an ROI. In this case it provides a means of detecting image offsets very quickly, unlike the movie technique, which requires time to set up the sequence of images. Rotations are extremely difficult to detect with this method. The process of fading between images has been found to be useful when comparing megavoltage and simulator images, particularly as the simulator images contain information beyond the field edge, whereas the megavoltage images do not. Summary and conclusions
In this paper we have presented an evaluation of the RMH/ICR megavoltage imaging system for patient position reproducibility studies. Various key considerations pertinent to the use of the system to make measurements have been presented: the need to have access to simulator images; the problem of finding and matching field edges; the delineation of soft-tissue interfaces; the efficacy of using regions of interest; and the desirable properties of a quantitative measurement procedure. Several algorithms for image measurement were presented. Four provide image intercomparison information: feature transposition, ROI transposition, movie measurement, and difference images. Three are qualitative: movie, side-by-side, and fading between images. A ruler, angle measurement tool was also discussed. The evaluation of the intercomparison algorithms shows that movie measurement, using the whole image, was the best method, followed by difference images, followed by ROI transposition, typically using the edge of a structure, followed by feature transposition, using a set of points. Thus, we conclude that the usefulness of these algorithms is determined by how much of the information content of the images is used. The difference image technique is useful for determining shifts in l. 65, No. 776
megavoltage images, but requires much time to implement (typically several minutes compared with 1 min or less for the other techniques). It also requires significant user familiarity because of its non-standard presentation of the image data. In conclusion, our megavoltage imaging system is intended for patient position verification studies. A short image acquisition time (
