Continuous monitoring of prostate position using stereoscopic and monoscopic kVimage guidanceM. Tynan R. Stevens, Dave D. Parsons, and James L. Robar Citation: Medical Physics 43, 2558 (2016); doi: 10.1118/1.4947295 View online: http://dx.doi.org/10.1118/1.4947295 View Table of Contents: http://scitation.aip.org/content/aapm/journal/medphys/43/5?ver=pdfcov Published by the American Association of Physicists in Medicine Articles you may be interested in The first clinical treatment with kilovoltage intrafraction monitoring (KIM): A real-time image guidance method Med. Phys. 42, 354 (2015); 10.1118/1.4904023 Evaluation of the geometric accuracy of surrogate-based gated VMAT using intrafraction kilovoltage x-rayimages Med. Phys. 39, 2686 (2012); 10.1118/1.4704729 Clinical development of a failure detection-based online repositioning strategy for prostateIMRT—Experiments, simulation, and dosimetry study Med. Phys. 37, 5287 (2010); 10.1118/1.3488887 Dosimetric consequences of misalignment and realignment in prostate 3DCRT using intramodality ultrasoundimage guidance Med. Phys. 37, 2787 (2010); 10.1118/1.3429127 Prostate intrafraction motion evaluation using kV fluoroscopy during treatment delivery: A feasibility andaccuracy study Med. Phys. 35, 1793 (2008); 10.1118/1.2899998

Continuous monitoring of prostate position using stereoscopicand monoscopic kV image guidance

M. Tynan R. Stevens, Dave D. Parsons, and James L. RobarDepartment of Medical Physics, Dalhousie University, Halifax, Nova Scotia B3H 4R2, Canada and NovaScotia Cancer Centre, QEII Health Science Centre, Halifax, Nova Scotia B3H 2Y9, Canada

(Received 18 November 2015; revised 29 March 2016; accepted for publication 9 April 2016;published 27 April 2016)

Purpose: To demonstrate continuous kV x-ray monitoring of prostate motion using both stereoscopicand monoscopic localizations, assess the spatial accuracy of these techniques, and evaluate the dosedelivered from the added image guidance.Methods: The authors implemented both stereoscopic and monoscopic fiducial localizations usinga room-mounted dual oblique x-ray system. Recently developed monoscopic 3D position estimationtechniques potentially overcome the issue of treatment head interference with stereoscopic imagingat certain gantry angles. To demonstrate continuous position monitoring, a gold fiducial marker wasplaced in an anthropomorphic phantom and placed on the Linac couch. The couch was used as aprogrammable translation stage. The couch was programmed with a series of patient prostate motiontrajectories exemplifying five distinct categories: stable prostate, slow drift, persistent excursion,transient excursion, and high frequency excursions. The phantom and fiducial were imaged using140 kVp, 0.63 mAs per image at 1 Hz for a 60 s monitoring period. Both stereoscopic and monoscopic3D localization accuracies were assessed by comparison to the ground-truth obtained from the Linaclog file. Imaging dose was also assessed, using optically stimulated luminescence dosimeter insertsin the phantom.Results: Stereoscopic localization accuracy varied between 0.13±0.05 and 0.33±0.30 mm,depending on the motion trajectory. Monoscopic localization accuracy varied from 0.2±0.1 to1.1±0.7 mm. The largest localization errors were typically observed in the left–right direction. Therewere significant differences in accuracy between the two monoscopic views, but which view wasbetter varied from trajectory to trajectory. The imaging dose was measured to be between 2 and15 µGy/mAs, depending on location in the phantom.Conclusions: The authors have demonstrated the first use of monoscopic localization for a room-mounted dual x-ray system. Three-dimensional position estimation from monoscopic imaging per-mits continuous, uninterrupted intrafraction motion monitoring even in the presence of gantryrotation, which may block kV sources or imagers. This potentially allows for more accurate treatmentdelivery, by ensuring that the prostate does not deviate substantially from the initial setup position.C 2016 American Association of Physicists in Medicine. [http://dx.doi.org/10.1118/1.4947295]

Key words: stereoscopic, monoscopic, intrafraction motion, prostate cancer, x-ray imaging

1. INTRODUCTION

The accurate delivery of external beam radiation therapydepends on precise localization of the anatomy to be irradiated.While planning CT and pretreatment imaging are routinelyused for patient setup, this cannot account for intrafractionmotion observed for many internal organs like the prostate.Indeed, prostate motion of more than 1 cm is not uncommon,and for most patients, the prostate will spend at least 5% ofthe treatment fraction more than 4 mm from the expectedlocation.1 These deviations from the setup position can affectthe dosimetric outcomes of treatment, as it has been shownthat motion of 5 mm can result in a 10% reduction of the 100%dose coverage.2 Although prostate motion is not always largeenough to produce serious dosimetric impact, intrafractionmotion monitoring can help avoid this possibility altogether.

An additional consideration regarding intrafraction motionis the impact of this positional uncertainty on planning target

volume (PTV) margins. PTV margins must be made largeenough to ensure that the prescribed dose to the clinical targetvolume (CTV) is maintained despite systematic and randomvariations in treatment delivery.3 However, smaller PTVmargins are desirable in order to reduce the dose delivered tohealthy tissue, particularly for nearby organs-at-risk (OARs).The widely used formula of van Herk et al.3 for calculatingPTV margins contains terms for both systematic preparationerrors and random variations throughout treatment. Thus,knowledge of intrafraction motion can be used to reduce PTVmargins by accounting for a substantial component of randomvariance,4 for example, by gating treatment or dynamicallyupdating couch5,6 or MLC positions.7–9

The impact of intrafraction motion is especially importantto consider given the recent interest in hypofractionatedradiotherapy of prostate tumors.10,11 Hypofractionation is anattractive option for prostate tumors because the generallyaccepted α/β ratio of prostate tumors is low compared to

2558 Med. Phys. 43 (5), May 2016 0094-2405/2016/43(5)/2558/11/$30.00 © 2016 Am. Assoc. Phys. Med. 2558

2559 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2559

the surrounding organs-at-risk,12 which allows for improvedtumor cell kill with potentially fewer adverse effects. Whereasit is typically assumed that position deviations average outacross fractions (i.e., contribute to random variation only),3

this assumption is invalid in hypofractionated treatment due tothe small number of fractions. It is thus especially important toemploy motion monitoring in hypofractionated treatment, inorder to achieve the desired tumor control and OAR sparing.

There are several techniques available for intrafraction mo-tion monitoring, including implanted RF transponders,7,13,14

stereoscopic x-ray imaging,15–17 or monoscopic imaging.18–22

Stereoscopic techniques include kV/MV imaging using theon-board imager (OBI) and MV beam’s eye view,17 and room-mounted dual16 or quad23,24 kV imaging. These techniquesusually rely on implanted gold fiducial markers as prostatesurrogates. Stereoscopic kV/MV imaging has the advantage ofwidespread availability; however, the fiducial markers can beobstructed in the MV images by movements of the MLC leaves,and the changing view-angle can result in variable fiducialoverlap with other fiducials or bony anatomy. Room-mountedkV systems are not affected by MLC positions and providea constant view angle, which makes it easier to ensure nooverlap of fiducials. However, dual kV room-mounted systemsfrequently have one of their tube/detector pairs blocked by thegantry at certain angles as it rotates around the patient (Fig. 1).While this shortcoming is addressed by quad kV systems likethat used by Shimizu et al.,23 these systems have not beenwidely adopted.

The intermittent blocking of one x-ray tube/detector bythe treatment head has presented a substantial challenge in theimplementation of intrafraction motion monitoring with room-mounted systems, as 3D location cannot be exactly determinedfrom a single 2D image perspective. However, techniques formonoscopic localization of fiducial markers have recentlybeen developed for OBI systems.19,22 While monoscopicimaging only provides absolute localization in two dimensions,

F. 1. Gantry angle restrictions for our room-mounted stereoscopic imagingsystem. When the gantry angle is in the red-shaded zones, the treatment headblocks one of the two stereo x-ray panels (1 and 2) or sources (3 and 4). Forthe green-shaded angles, both panels have unobstructed views of isocenter.The exact angles available for stereo imaging will vary from setup to setup.

correlations between prostate motion in the anterior–posteriorand inferior–superior directions can be used in order to performan informed estimate of the unresolved dimension. Although3D position estimation from monoscopic imaging is naturallyless accurate than stereoscopic localization, it can achieve sub-mm accuracy,19 and therefore is a substantial improvement onno intrafraction monitoring.

While monoscopic localization has been demonstrated forOBI systems, to our knowledge no studies have investigated theuse of monoscopic localization for room-mounted kV systems.We therefore aim to demonstrate monoscopic localizationusing a room-mounted kV system. This technique alleviatesthe issue of restricted gantry angles, enabling uninterruptedintrafraction motion monitoring for room-mounted systems. Inthis work, we demonstrate the accuracy of the monoscopic andstereoscopic localization technique in phantom by comparisonwith a ground-truth trajectory and assess the extra dosedelivered by the added kV imaging.

2. METHODS

We evaluated monoscopic localization accuracy using aroom-mounted dual kV imaging system. We identified threemotion monitoring schemes: (1) full stereoscopic localization,(2) monoscopic localization (using either x-ray sources, seeFig. 1), and (3) no imaging. In all cases, the gantry wasparked at zero degrees to prevent image obstruction, andmonoscopic image series were created retrospectively. Theno imaging scenario represents a worst case, as only the initialsetup position is known. For each scheme, we evaluated theaccuracy by comparison with prescribed phantom motion anddetermined the extra imaging dose delivered for the purposeof motion monitoring.

2.A. Motion phantom

A cylindrical gold fiducial marker (1× 5 mm) was fixedbetween two layers of an anthropomorphic phantom (ATOMDosimetry Phantoms, Norfolk, VA) at the approximate loca-tion of the prostate gland (Fig. 2). The phantom was placedon the treatment couch of the linear accelerator (Varian STx,Varian Medical Systems, Inc.), and the alignment lasers wereused to place the fiducial marker initially at isocenter. TheLinac couch was utilized as a programmable translation stageusing Developer Mode, with real prostate motion trajectoriesimplemented from published Calypso-based patient data.21

Five unique trajectories of approximately 1 min durationwere implemented, including stable prostate position, slowdrift, transient excursion, persistent excursion, and frequentexcursions.21,22 For each motion trajectory, a log file of thecouch positions produced by the Linac was collected as theground-truth locations.

2.B. Image acquisition and analysis

Monoscopic and stereoscopic x-ray imaging was performedusing a room-mounted dual kV system (ExacTrac, Brainlab

Medical Physics, Vol. 43, No. 5, May 2016

2560 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2560

F. 2. Experimental setup: (a) A fiducial marker (x) was fixed between two slabs of the anthropomorphic phantom. The phantom also contained several insertsfor optically stimulated luminescent dosimeters. (b) The phantom was placed on the Linac couch, which was used as a programmable translation stage, andthe fiducial motion was imaged using the room-mounted x-ray system. The couch motion and image data were temporally coregistered using a microcomputerequipped with an accelerometer and two field diodes.

AG, Feldkirchen, Germany). This system consists of two x-ray sources embedded in the floor of the treatment room,and two ceiling mounted flat panel detectors, providingorthogonal oblique image projections with approximately13.5 cm field of view at isocenter (Fig. 3). Continuous imageswere acquired at a rate of 1 Hz throughout each motiontrajectory, resulting in 60 images per experiment. The x-ray tubes were set to deliver 0.63 mAs per acquisition ata tube potential of 140 kVp, to provide adequate fiducialcontrast.

The fiducial markers were automatically detected from thex-ray images using a maximum convolution approach.25,26 Forthis purpose, a 64×64 pixel convolution kernel was createdwith the central pixels set to 1, surrounded by a 1 pixelborder with a negative value, determined such that the summedpixel value of the entire kernel is zero. When convolved withan image, this kernel thus produces zero for features muchlarger than the central kernel area, and maximal values forfeatures the same size and orientation as this central region.To optimize the detection process, the angle and size of thefiducial marker in the two projections were determined froman initial image acquisition pair. As the acquisition geometryis constant for room-mounted systems, this projected shape isconsistent from image to image. For each image projection, the

point of maximum convolution was taken as the image locationof the fiducial marker (i.e., [i1, j1] and [i2, j2] in Fig. 3). From thedetected image locations of the fiducials, patient coordinatesfor the fiducial markers were computed via both stereoscopicand monoscopic approaches (see the Appendix for detailedcalculations).

In order to relate the fiducial locations determined byimaging to the known couch positions obtained from theLinac log file, a common temporal frame of reference isrequired. A custom microcomputer was built for this purpose(Fig. 2), which included an accelerometer to detect the couchmotion, and two field diode inputs to monitor the x-ray output.Continuous recording from each of these three devices wasperformed throughout each motion monitoring experiment.With this system, the exact timing of image acquisitionwith respect to the motion trajectories was determined bycoregistering the onset of movement in the accelerometerdata to the Linac log file. The localization accuracy wasassessed by the root-mean-square (RMS) localization er-ror (averaged across all imaging time-points), using theLinac log file positions as ground-truth. Significance ofaccuracy differences was assessed using a paired t-test ofthe accuracy versus time for each localization method andtrajectory.

Medical Physics, Vol. 43, No. 5, May 2016

2561 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2561

F. 3. The x-ray imaging geometry is specified by the spherical coordinatesof the image sources and detectors, parameterized by the polar angle “Θ,”azimuthal angle “Φ,” source to axis (isocenter) distance (SAD), and axis todetector distance (ADD). The image coordinate systems can be specified interms of pixel locations ([i1, j1] and [i2, j2]) or mm relative to the projectionof isocenter ([xim1, yim1] and [xim2, yim2], respectively). The above geometryis then used to convert image locations to patient or room coordinates ([xp,yp, zp]).

2.C. Measurement of imaging dose

We assessed the dose associated with the additional kV im-ages required for motion monitoring, as any additional patientdose represents an important consideration when adopting thistechnique. For this purpose, optically stimulated luminescentdosimeters (OSLDs, Landauer Nanodots, Chicago, IL) wereplaced in prefabricated inserts at various locations in thephantom slabs immediately above and below the fiducialmarker [Fig. 2(a)]. The OSLDs were cross-calibrated againstan ion chamber (Exradin A12, Standard Imaging, Middleton,WI) calibrated for the same kVp and half-value layer followingTG-61. To achieve sufficient signal on the OSLDs, the phantomwas exposed to 50 acquisitions from a single x-ray tubeoperating at 140 kVp and 40 mAs. The dose from the othertube and from stereoscopic acquisitions was then inferred bysymmetry.

2.D. MV scatter

The experiments described above were performed with theMV treatment beam off, in order to exclude the influence ofMV scatter on accuracy of fiducial detection in the images, andsubsequently 3D localization. To ensure that the methods usedare viable in situ, we evaluated reproducibility of the fiducialdetection with the treatment beam on, using a 5×5 cm2 fieldand a dose rate of 500 MU/min. We tested the two worst-case-scenarios for MV scatter: gantry at 180◦ (when the beam ismost pointed toward the detectors) and 90◦ (where the lateraldimension of the patient creates the most scatter). For eachgantry angle, we collected 20 images at four different kV mAs

settings (0.63, 1.0, 1.2, and 1.6 mAs). The variance in thedetected fiducial location across the 20 images was assessedfor each mAs and gantry angle pairing.

3. RESULTS3.A. Localization accuracy

The results of motion monitoring using stereoscopic ormonoscopic imaging are shown in Fig. 4 for the variousclasses of prostate motion. Stereoscopic imaging producedvery accurate localization, with less than 0.4 mm RMS errorfor all five trajectories (Fig. 5). The highest RMS error for thestereoscopic localization was for the “persistent excursion”trajectory (0.34 ± 0.30 mm), for which there was a brieflocalization error around the time of the excursion. Thiserror was the largest in the left–right direction, with a peak3D mislocalization of 1.8 mm. In the trajectories with rapidprostate motion (e.g., frequent excursions), the imaging rateof 1 Hz can be insufficient to sample the full dynamics. Whilethis typically only results in small interpolation errors, higherimaging frequency may be desirable.

Monoscopic localization also produced sub-mm accuracyon average in 3D, in all but one case (tube 1, transientexcursion trajectory). In this case, the RMS error was onlymarginally larger than 1 mm (1.1 ± 0.7 mm), largely due toan underestimation of the excursion amplitude around t = 41 s[Fig. 4(d)], and an overestimation of motion in the left–rightdirection at the same time. A similar underestimation ofmotion in the ant/post and sup/inf directions, accompaniedby a mislocalization in the left/right direction was observedusing tube 2 in the slow drift trajectory [Fig. 4(b)]. Thelargest error in monoscopic localization was approximately4 mm (tube 1, transient excursion at 41 s), but still accountedfor the majority of the relatively large (∼12 mm) and rapid(∼5 s) excursion at that point in the trajectory. In general,monoscopic imaging accurately detected most of the motionin all trajectories.

For all trajectories, stereoscopic imaging produced signif-icantly better localization accuracy than no imaging (p< 0.001). For monoscopic imaging, all but the stable prostatetrajectory produced significantly better localization than noimaging (p < 0.001), with no difference in the case of the stableprostate. Stereoscopic localization was significantly moreaccurate than monoscopic at the p < 0.001 level except forthe persistent excursion trajectory, which was only significantat p < 0.05 for monoscopic tube 2. While there were significantdifferences in the accuracy of the two views for monoscopiclocalization, the superior view varied between trajectories.

3.B. Imaging dose

The imaging dose per mAs delivered by a single x-ray tube is shown in Fig. 6 for a number of locations inthe slab above and below the fiducial marker. The peakdose (15.3± 0.1 µGy/mAs) was observed at the posteriorof the phantom, nearest to the beam entry point. Dose wassignificantly higher (p < 0.001) in the slab below the fiducial

Medical Physics, Vol. 43, No. 5, May 2016

2562 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2562

F. 4. Actual and reconstructed fiducial marker trajectories for a variety of prostate trajectories. Examples of (top to bottom) stable prostate, slow drift, persistentexcursions, transient excursions, and high frequency excursions are shown.

(i.e., closer to the gantry), as the beams enter from this direction(see Fig. 3). The oblique orientation of the x-ray sourcesalso causes the strong dose gradient in the anterior/posteriordirection shown in Fig. 6.

The protocol in this study used 0.63 mAs per image,resulting in 37% lower dose per image than what is shownin Fig. 6. This was the lowest mAs setting available onour x-ray system and produced sufficient image quality toidentify the fiducial markers. During stereoscopic imaging,when both x-ray tubes are firing, the dose will be equivalentto about 1.26 mAs per image pair. The total dose delivered isproportional to the number of images acquired, which itself isthe product of imaging frequency and fraction duration. Using1 Hz imaging as in this study, a 3 min VMAT plan would resultin a peak dose of 5.4 ± 0.1 mGy in the case of stereoscopicimaging (worst-case-scenario).

3.C. MV scatter

With the MV beam on, it was necessary to increasethe kV tube current in order to clearly resolve the fiducialin the images (Fig. 7). For kV tube current of 1.2 mAsor greater, the fiducial was detected in the images withperfect reproducibility. Although difficult to see in the rawimages, scatter was more problematic with the gantry at90◦ than 180◦, due to the thicker lateral dimension of thephantom.

4. DISCUSSION

To our knowledge, this is the first report of monoscopicmotion monitoring, with estimate of 3D position, using aroom-mounted kV x-ray system. This avoids the most serious

Medical Physics, Vol. 43, No. 5, May 2016

2563 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2563

F. 5. Accuracy of the motion monitoring techniques. The 25th to 75th per-centiles are indicated by the boxes, with the median error shown by the whitelines. The whiskers end at the 5th and 95th percentiles, with outliers shownwith circles. For the slowly moving prostates (stable, drift, and persistentexcursions), the no imaging case results in large median errors, whereas forthe prostates with rapid dynamics (transient and frequent excursions), thelargest deviations appear as outliers. All three motion monitoring techniquesaccurately account for the majority of the prostate motion.

pitfall of motion monitoring using room-mounted systems,i.e., blocking of the x-ray tubes by the Linac treatmenthead. Both stereoscopic motion and monoscopic motionmonitoring successfully reduced the intrafraction positionaluncertainty for this simulated prostate treatment. Although bydefinition stereoscopic localization is capable of producingmore accurate results, monoscopic localization was able toaccount for the majority of prostate motion, significantlyreducing position uncertainty. The knowledge of actual pros-tate position throughout treatment can be used as feed-forwardto dynamic MLC (Refs. 7–9) or couch position updates5,6

in order to increase the accuracy and conformality of dosedelivery.

Our monoscopic error localization errors ranged from 0.2to 1.1 mm RMS, with a maximum of 3.8 mm. For comparison,stereoscopic localization errors ranged from 0.1 to 0.3 mmRMS, with a maximum of 1.8 mm. Previous studies usingthe same monoscopic localization algorithm for OBI imagingsystems found approximately 0.35 mm RMS error, with a3.6 mm maximum error.20 Using a dynamically updated prob-ability density function (PDF), slightly smaller RMS error was

achievable (0.22 mm),20 depending on the prostate trajectory.A study by Poulsen et al.21 using the same prostate trajectoriesas in our study found a mean RMS error of 0.3–1.0 mm, witha maximum localization error of 2 mm. The main competitivetechnique is implanted electromagnetic transponders, withreported localization accuracy of 0.4–1.5 mm.1,7,13 Overall,the monoscopic localization accuracy of this study com-pares favorably with those reported previously in the litera-ture, and the stereoscopic accuracy exceeds other availablemethods.

The potential impact of the intrafraction monitoring on PTVmargins (MPTV) can be estimated using the formula of van Herket al.:3 MPTV= 2.5Σ+0.7σ, where Σ and σ are the systematicand random setup variances respectively, including both inter-fraction and intrafraction errors (e.g., σ =

√(σ2inter+σ

2intra)).

The potential margin reductions (∆M) afforded by resolvingintrafraction motion are thus ∆M = 0.7(σ−σmonitored), whereσmonitored includes interfraction random error, as well as theresidual variance unaccounted for by intrafraction motionmonitoring (i.e., σmonitored =

√(σ2inter+σ

2residual)). Taking the

value of σinter = (1.4 mm LR, 1.6 mm AP, 1.4 mm SI)for initial setup using fiducial marker surrogates from Tanyiet al.,4 the PTV margin reduction made possible by accountingfor intrafraction motion ranges from negligible in the stableprostate case to (0, 1.5, 0.6 mm) in the case of the persistentexcursion. For the four relatively dynamic prostate trajectoriesinvestigated, there was no difference between the marginreduction for monoscopic (0, 1.1, 0.5 mm) and stereoscopic (0,1.2, 0.5 mm) motion monitoring on average. Importantly, thelargest margin reduction was observed in the AP direction,which is the direction of one of the principal OARs forprostate treatment—the rectal wall. These margin reductionswould need to be facilitated by either restrictive motiongating tolerances or (preferably) highly accurate trackingmechanisms, such as the dynamic MLC system proposed byPoulsen et al.8

There are several inherent advantages of the room-mountedx-ray system compared to the OBI. With the room-mountedsystem, a pretreatment imaging period (i.e., with the gantryparked at vertical) can be used to build a patient-specific PDFunder the same geometry as will be used for monitoring.This direct measurement of motion variance/covariance ismore efficient than the alternative probabilistic techniquesneeded in the OBI case.20 Furthermore, there is also no sagor flex corrections needed to compensate for the change inisocenter/imager relationships as the gantry rotates. Finally,the constant view angle simplifies the problem of detectingthe fiducial markers in the images, as there is no change inthe projection geometry, and likewise no substantial changesin fiducial overlap with other fiducials or with backgroundanatomy. While both of these fiducial detection issues can beovercome via sophisticated detection strategies [e.g., see thework of Feldelius et al. 2011 (Ref. 27) and 2014 (Ref. 28)], theadded complexities will reduce overall sensitivity/specificityof detection.

An additional consideration regarding imaging geometry isthat not all motions are equally important from a dosimetricperspective. Due to the rapid fall off of typical beam profiles,

Medical Physics, Vol. 43, No. 5, May 2016

2564 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2564

F. 6. Dose per mAs at the locations of the OSLD dosimeters in the phantom slabs above (a) and below (b) the fiducial marker. Peak dose was observed at theposterior aspect of the phantom in the slice below the fiducial marker (i.e., closer to the gantry), due to the beam entry angle.

motion perpendicular to the beam’s-eye-view (BEV) is moredetrimental to target coverage than motion along the BEV.29,30

In the case of OBI imaging, motion along the BEV is alwaysresolved directly, whereas one of the axes perpendicular toBEV is not directly observable. In the case of the room-mounted geometry, the monoscopic image plane is at an angleto the BEV, such that both perpendicular axes are partiallyresolved. Alternatively, in-line or BEV imaging in theory iscapable of resolving all off-axis motion directly, but suffersfrom MLC blockage and poor MV image contrast. A poten-tially useful development is in-line kilovoltage imaging;31,32

however, in order to achieve the desired energy spectrum,special Linac targets must be used, preventing this techniquefrom being employed during treatment. We are currentlyinvestigating a rapidly switching target system to overcomethis issue,33 with one of the potential applications beingintrafraction imaging.

It is worth pointing out that for each trajectory, thelocalization method with the highest error represents aworst-case-scenario. In practice, gantry rotation during thetreatment fraction would necessitate toggling the tube usedfor monoscopic imaging as the treatment head rotates (i.e.,quadrants 1 and 3 versus 2 and 4 in Fig. 1), and indeedwould allow for periods of stereoscopic monitoring aroundeach of the cardinal angles. Thus the accuracy obtainedin a realistic implementation would be somewhere in themiddle of the range shown in Fig. 6. The added informationafforded by the intermittent stereoscopic imaging opportu-nities could potentially be used to improve the monoscopiclocalization model and should be investigated in futurestudies.

In general, the monoscopic localization technique washighly accurate in the sup/inf and ant/post directions, withthe largest errors occurring in the lateral direction. This resultis likely due to a combination of two factors: (1) motionin the left/right directions is much less correlated with theother two dimensions and is thus less certain to infer from the2D measurement, and (2) the geometry of the room-mountedimaging system is such that (unlike the lab-frame covariancematrix), the rotated covariance matrix has no near-zero entries.Because of this latter point, variance in the sup/inf and ant/postdirections tends to get projected somewhat into the lateraldimension.

The added dose to patients of approximately 5 mGy perfraction is low in the context of a typical 2 Gy fraction. Overthe course of a typically fractionated treatment, this could addup to as much as 20 cGy, or 10% of a single fraction dose,assuming conventional fractionation. This result is comparableto previous estimates of 10–15 cGy from 1 Hz imaging over thecourse of a VMAT treatment using OBI imaging.34 The onlypublished data on ExacTrac imaging dose available estimate0.551 mGy entrance dose per image at 140 kVp and 15 mAs,35

or equivalently 37 µGy/mAs. This is about 2.5 times ourobserved dose at the depth of the rectum, which is consistentwith the PDD at this depth and energy.36

Any increase in the image rate or treatment time changesto the imaging technique (kVp or mAs), or added pretreat-ment imaging can increase the imaging dose. However, inhypofractionated treatment, the total treatment time is typicallyreduced, which lowers the relative contribution of imagingdose for a fixed image frequency.34 Thus hypofractionatedregimens are not only the best candidate for prostate motion

Medical Physics, Vol. 43, No. 5, May 2016

2565 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2565

F. 7. Reproducibility of fiducial detection in the presence of MV scatter, for gantry angles of (a) 90◦ and (b) 180◦. A small ROI of representative raw imagesaround the fiducial marker is shown in the top rows of (a) and (b). The mean ± standard deviation of the detected fiducial location (xi, yi) was determined from20 replications of the images at each of four tube current levels (bottom rows). At 0.63 mAs, the fiducials are obscured by the MV scatter contribution, resultingin unreliable fiducial detection. For 1.2 mAs or higher, the fiducial is reliably detected at both gantry angles.

monitoring, but benefit the most from the added localizationcertainty.

In this study, we initially ignored the influence of MVscatter on image quality, and subsequently on the accuracyof fiducial detection and localization. It has previously beenshown for OBI imaging that accurate fiducial detection isachievable even at high MV dose rates,27,28 and the increasedisocenter-to-detector distance of the ceiling mounted panelsshould further reduce the added scatter noise. Nonetheless,the presence of MV scatter noise necessitated increasedmAs to maintain adequate image quality. Our results suggestthat tube current of approximately 1.2 mAs would producesufficient image quality for typical MV dose rates and field

sizes. While this would approximately double the imagingdose, the result would still only represent 0.5% of a typicaltreatment fraction and could in principle be accounted forduring treatment planning. Additionally, the dose estimatesabove were produced using the conservative assumption ofconstant stereoscopic imaging, which in practice would bereduced by almost half given the large range of gantry anglesfor which monoscopic imaging would be used. A moresophisticated approach would be to modulate the kV imagingtechnique as a function of gantry angle and (MLC-modulated)field size in order to provide constant image quality whileminimizing accumulated dose, but this is outside the scope ofthis paper.

Medical Physics, Vol. 43, No. 5, May 2016

2566 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2566

F. 8. Illustration of (a) stereoscopic and (b) monoscopic localizationschemes employed (shown in 2D for simplicity of illustration). Stereoscopiclocalization finds the intersection point (black circle) of the ray lines con-necting the detected image locations (i1,i2) of the fiducials (grey filledcircles) with the corresponding source locations (s1 and s2, respectively).Monoscopic localization finds the maximum likelihood position (black cir-cle) along the single ray line connecting the detected image location andsource point using a PDF derived from the motion covariances (Px, y).Monolocalization and stereolocalization modes can be toggled in real timedepending on which imaging panels have an unobstructed view of the fiducialmarkers.

5. CONCLUSION

We have implemented both stereoscopic and monoscopicmotion monitorings using a room-mounted kV imagingsystem. The availability of both localization modes allowscontinuous 3D localization despite the wide range of anglesover which the treatment head may obstruct one of the x-ray tubes and detectors. The motion monitoring accuracyrivals that available with other published methods and issuperior when stereoscopic views are available. Accurateintrafraction prostate localization is especially beneficial forhypofractionated treatment where movement of the targetvolume can have critical dosimetric impact.

ACKNOWLEDGMENTS

The authors would like to acknowledge financial supportfrom Varian Medical Systems and technical support fromBrainlab AG. Lee MacDonald provided guidance on usingdeveloper mode for driving motion trajectories and Dr.Mike Sattarivand who contributed details on the geometryof their specific ExacTrac system setup. The authors alsorecognize valuable contribution from machinist John Noddinfor creating the custom OSLD inserts for their phantomexperiments.

APPENDIX: LOCALIZATION CALCULATIONS1. Stereoscopic localization

To find patient coordinates (r= [xp,yp,zp]) of the fiducialmarker using stereoscopic localization (rstereo), the method ofBrost et al.15 is followed. Using the geometry shown in Fig. 3,the image locations (p1= [i1, j1,0] and p2= [i2, j2,0], in pixels)are first converted to room coordinates (d1 = [xd1,yd1,zd1]

and d2= [xd2,yd2,zd2] respectively). For example, for the firstdetector,

d1=R · [(p1−po1)◦dp]+do1=R ·

xim1

yim1

0

+do1, (A1)

where R is the rotation matrix that aligns unit vectors inthe detector plane with the room coordinate system (seeFig. 3), po1= [io1, jo1,0] is the pixel coordinates of the projectedisocenter, dp = [dx,dy,0] is the vector of pixel dimensions,and do1 = [xo1,yo1,zo1] are the room coordinate for the pro-jected isocenter (i.e., ADD∗[cos Θ cos Φ, cos Θ sin Φ, cos Θ]).Points (r1 = [xp1,y p1,zp1]) on the line connecting the de-tected point with the source location (s1 = [xs1,y s1,zs1]= SAD∗[−cos Θ cos Φ,−cos Θ sin Φ,−cos Θ]) can then beparameterized as

r1=d1+m1(s1−d1)=d1+m1a1, (A2)

where m1 governs the distance along the ray line from thedetector. Likewise for the line connecting the second detectorand source,

r2=d2+m2a2. (A3)

The location of the fiducial marker is the point at which thesetwo lines intersect (Fig. 8), and thus

d1+m1a1=d2+m2a2 (A4)

or equivalently

d=A ·

m1

−m2

, (A5)

where d=d2−d1 and A= [a1,a2]. In practice the two ray lineswill not intersect perfectly due to finite measurement precision;however, an approximate solution to this equation is given byfinding the pseudo-inverse for A

�A†

�,

A†=�AT ·A

�−1 ·AT, (A6)

yielding

m1

−m2

=A†d. (A7)

The values of m1 and m2 are then used to estimate the two(ideally equivalent) world points (r1 and r2), and the mean ofthese is used as the final stereoscopic localization rstereo.

2. Monoscopic localization

The algorithm employed for 3D localization from mono-scopic imaging relies on a priori information in the formof motion covariances “C” (Fig. 8), which can be obtaineddirectly from stereoscopic localization pretreatment, or from

Medical Physics, Vol. 43, No. 5, May 2016

2567 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2567

published population averages,19

C =

varx covxy covxz

covxy vary covyz

covxz covyz varz

=

0.3163 −0.0775 0.0114−0.0775 2.4733 1.50510.0114 1.5051 1.8820

mm2. (A8)

This covariance matrix can be used to determine a GaussianPDF for the fiducial locations (P(x,y,z)),

P(x,y,z)=

det(C−1)8π3 e−rTC−1r/2 (A9)

or in the coordinate system rotated to be parallel to the imageplane,

Prot(xrot,yrot,zrot)=

det(C−1)8π3 e−rT

rotR−1C−1Rrrot/2. (A10)

Identifying the matrix elements of the rotated covariance as

R−1C−1R=

Arot Drot/2 Erot/2Drot/2 Brot Frot/2Erot/2 Frot/2 Crot

. (A11)

Poulsen et al.19 showed that the expectation value for theposition along the axis perpendicular to the imaging plane(⟨zrot⟩) is

⟨zrot⟩ = SADArot

( xim

SDD

)2+Brot

(yim

SDD

)2+Drot

ximyim

SDD2

− Erotxim

2 ·SDD−Frot

yim

2 ·SDD

σ2, (A12)

where SDD is the source-to-detector distance (i.e., SDD =SAD+ADD) and σ is the standard deviation,

σ =

Arot

( xim

SDD

)2+Brot

(yim

SDD

)2+Crot

+ Drotximyim

SDD2 −Erotxim

SDD−Frot

yim

SDD

−1/2

. (A13)

By appropriately scaling the image coordinates ([xim,yim])according to zrot, we obtain the (rotated) 3D location of thefiducial marker (rrot= [xrot,yrot,zrot]),

xrot= xim(SAD− zrot)/(SAD+ADD), (A14)yrot= yim(SAD− zrot)/(SAD+ADD). (A15)

Finally, the monoscopic localization coordinates (rmono) areobtained by applying the rotation matrix (i.e., rmono=R ·rrot).

1H. S. Li, I. J. Chetty, C. A. Enke, R. D. Foster, T. R. Willoughby, P. A.Kupellian, and T. D. Solberg, “Dosimetric consequences of intrafractionprostate motion,” Int. J. Radiat. Oncol. 71(3), 801–812 (2008).

2S. Hossain, P. Xia, C. Chuang, L. Verhey, A. R. Gottschalk, G. Mu, and L.Ma, “Simulated real time image guided intrafraction tracking-delivery forhypofractionated prostate IMRT,” Med. Phys. 35(9), 4041–4048 (2008).

3M. van Herk, P. Remeijer, C. Rasch, and J. V. Lebesque, “The probabilityof correct target dosage: Dose-population histograms for deriving treatment

margins in radiotherapy,” Int. J. Radiat. Oncol. Biol. Phys. 47(4), 1121–1135(2000).

4J. A. Tanyi, T. He, P. A. Summers, R. G. Mburu, C. M. Kato, S. M. Rhodes,A. Y. Hung, and M. Fuss, “Assessment of planning target volume margins forintensity-modulated radiotherapy of the prostate gland: Role of daily inter-and intrafraction motion,” Int. J. Radiat. Oncol. 78(5), 1579–1585 (2010).

5W. D. D’Souza, S. A. Naqvi, and C. X. Yu, “Real-time intra-fraction-motiontracking using the treatment couch: A feasibility study,” Phys. Med. Biol.50(17), 4021–4033 (2005).

6P. Qiu, W. D. D’Souza, T. J. McAvoy, and K. J. Ray liu, “Inferential modelingand predictive feedback control in real-time motion compensation using thetreatment couch during radiotherapy,” Phys. Med. Biol. 52(19), 5831–5854(2007).

7A. Sawant, R. L. Smith, R. B. Venkat, L. Santanam, B. Cho, P. Poulsen,H. Cattell, L. J. Newell, P. Parikh, and P. J. Keall, “Toward submillimeteraccuracy in the management of intrafraction motion: The integration of real-time internal position monitoring and multileaf collimator target tracking,”Int. J. Radiat. Oncol. 74(2), 575–582 (2009).

8P. R. Poulsen, W. Fledelius, B. Cho, and P. Keall, “Image-based dynamicmultileaf collimator tracking of moving targets during Intensity-modulatedarc therapy,” Int. J. Radiat. Oncol. 83(2), e265–e271 (2012).

9P. J. Keall, H. Cattell, D. Pokhrel, S. Dieterich, K. H. Wong, M. J. Murphy,S. S. Vedam, K. Wijesooriya, and R. Mohan, “Geometric accuracy of areal-time target tracking system with dynamic multileaf collimator trackingsystem,” Int. J. Radiat. Oncol. 65(5), 1579–1584 (2006).

10W. R. Lee, “Prostate cancer and the hypofractionation hypothesis,” J. Clin.Oncol. 31(31), 3849–3851 (2013).

11N.-S. Hegemann, M. Guckenberger, C. Belka, U. Ganswindt, F. Manapov,and M. Li, “Hypofractionated radiotherapy for prostate cancer,” Radiat.Oncol. 9(1), 275–290 (2014).

12S. M. Bentzen and M. A. Ritter, “The α/β ratio for prostate cancer: What isit, really?,” Radiother. Oncol. 76(1), 1–3 (2005).

13T. R. Willoughby, P. A. Kupelian, J. Pouliot, K. Shinohara, M. Aubin, M.Roach, L. L. Skrumeda, J. M. Balter, D. W. Litzenberg, S. W. Hadley, J. T.Wei, and H. M. Sandler, “Target localization and real-time tracking using theCalypso 4D localization system in patients with localized prostate cancer,”Int. J. Radiat. Oncol. 65(2), 528–534 (2006).

14T. Mate, D. Krag, J. Wright, and S. Dimmer, “A new system to performcontinuous target tracking for radiation and surgery using non-ionizing alter-nating current electromagnetics,” Int. Congr. Ser. 1268, 425–430 (2004).

15A. Brost, N. Strobel, L. Yatziv, W. Gilson, B. Meyer, J. Hornegger, J. Lewin,and F. Wacker, “Accuracy of x-ray image-based 3D localization from twoC-arm views: A comparison between an ideal system and a real device,”Proc. SPIE 7261, 72611Z (2009).

16M. S. Hoogeman, J. J. Nuyttens, P. C. Levendag, and B. J. M. Heijmen,“Time dependence of intrafraction patient motion assessed by repeat stereo-scopic imaging,” Int. J. Radiat. Oncol. 70(2), 609–618 (2008).

17R. D. Wiersma, W. Mao, and L. Xing, “Combined kV and MV imagingfor real-time tracking of implanted fiducial markers,” Med. Phys. 35(4),1191–1198 (2008).

18J. Adamson and Q. Wu, “Prostate intrafraction motion evaluation using kVfluoroscopy during treatment delivery: A feasibility and accuracy study,”Med. Phys. 35(5), 1793–1806 (2008).

19P. R. Poulsen, B. Cho, K. Langen, P. Kupelian, and P. J. Keall, “Three-dimensional prostate position estimation with a single x-ray imager utiliz-ing the spatial probability density,” Phys. Med. Biol. 53(16), 4331–4353(2008).

20P. R. Poulsen, B. Cho, and P. J. Keall, “Real-time prostate trajectory esti-mation with a single imager in arc radiotherapy: A simulation study,” Phys.Med. Biol. 54(13), 4019–4035 (2009).

21P. R. Poulsen, B. Cho, A. Sawant, and P. J. Keall, “Implementation of a newmethod for dynamic multileaf collimator tracking of prostate motion in arcradiotherapy using a single kV imager,” Int. J. Radiat. Oncol. 76(3), 914–923(2010).

22J. A. Ng, J. T. Booth, P. R. Poulsen, W. Fledelius, E. S. Worm, T. Eade, F.Hegi, A. Kneebone, Z. Kuncic, and P. J. Keall, “Kilovoltage intrafractionmonitoring for prostate intensity modulated arc therapy: First clinical re-sults,” Int. J. Radiat. Oncol. 84(5), e655–e661 (2012).

23S. Shimizu, H. Shirato, K. Kitamura, N. Shinohara, T. Harabayashi, T.Tsukamoto, T. Koyanagi, and K. Miyasaka, “Use of an implanted markerand real-time tracking of the marker for the positioning of prostate andbladder cancers,” Int. J. Radiat. Oncol. Biol. Phys. 48(5), 1591–1597(2000).

Medical Physics, Vol. 43, No. 5, May 2016

2568 Stevens, Parsons, and Robar: Stereoscopic and monoscopic imaging for prostate motion monitoring 2568

24K. Kitamura, H. Shirato, Y. Seppenwoolde, T. Shimizu, Y. Kodama, H.Endo, R. Onimaru, M. Oda, K. Fujita, S. Shimizu, and K. Miyasaka, “Tu-mor location, cirrhosis, and surgical history contribute to tumor movementin the liver, as measured during stereotactic irradiation using a real-timetumor-tracking radiotherapy system,” Int. J. Radiat. Oncol. 56(1), 221–228(2003).

25A. Nederveen, J. Lagendijk, and P. Hofman, “Detection of fiducial goldmarkers for automatic on-line megavoltage position verification using amarker extraction kernel (MEK),” Int. J. Radiat. Oncol. Biol. Phys. 47(5),1435–1442 (2000).

26E. J. Harris, H. A. McNair, and P. M. Evans, “Feasibility of fully automateddetection of fiducial markers implanted into the prostate using electronicportal imaging: A comparison of methods,” Int. J. Radiat. Oncol. 66(4),1263–1270 (2006).

27W. Fledelius, E. Worm, U. V. Elstrøm, J. B. Petersen, C. Grau, M. Høyer,and P. R. Poulsen, “Robust automatic segmentation of multiple implantedcylindrical gold fiducial markers in cone-beam CT projections,” Med. Phys.38(12), 6351–6361 (2011).

28W. Fledelius, E. Worm, M. Høyer, C. Grau, and P. R. Poulsen, “Real-timesegmentation of multiple implanted cylindrical liver markers in kilovolt-age and megavoltage x-ray images,” Phys. Med. Biol. 59(11), 2787–2800(2014).

29S. Nill, J. Unkelbach, L. Dietrich, and U. Oelfke, “Online correction forrespiratory motion: Evaluation of two different imaging geometries,” Phys.Med. Biol. 50(17), 4087–4096 (2005).

30Y. Suh, S. Dieterich, and P. J. Keall, “Geometric uncertainty of 2D projec-tion imaging in monitoring 3D tumor motion,” Phys. Med. Biol. 52(12),3439–3454 (2007).

31Y. Dzierma, F. G. Nuesken, N. P. Licht, and C. Ruebe, “Dosimetric prop-erties and commissioning of cone-beam CT image beam line with a carbontarget,” Strahlenther. Onkol. 189(7), 566–572 (2013).

32J. Rottmann, P. Keall, and R. Berbeco, “Real-time soft tissue motion esti-mation for lung tumors during radiotherapy delivery,” Med. Phys. 40(9),091713 (7pp.) (2013).

33R. I. Berbeco, A. Detappe, P. Tsiamas, D. Parsons, M. Yewondwossen, andJ. Robar, “Low Z target switching to increase tumor endothelial cell doseenhancement during gold nanoparticle-aided radiation therapy,” Med. Phys.43(1), 436–442 (2016).

34J. K. Crocker, J. A. Ng, P. J. Keall, and J. T. Booth, “Measurement of patientimaging dose for real-time kilovoltage x-ray intrafraction tumour posi-tion monitoring in prostate patients,” Phys. Med. Biol. 57(10), 2969–2980(2012).

35M. J. Murphy, J. Balter, S. Balter, J. A. BenComo, I. J. Das, S. B. Jiang,C.-M. Ma, G. H. Olivera, R. F. Rodebaugh, K. J. Ruchala, H. Shirato,and F.-F. Yin, “The management of imaging dose during image-guidedradiotherapy: Report of the AAPM Task Group 75,” Med. Phys. 34(10),4041–4063 (2007).

36D. Parsons and J. L. Robar, “An investigation of kV CBCT image quality anddose reduction for volume-of-interest imaging using dynamic collimation,”Med. Phys. 42(9), 5258–5269 (2015).

Medical Physics, Vol. 43, No. 5, May 2016