Sensitivity study of an automated system for daily patient QA using EPID exit dose images

Abstract The dosimetric consequences of errors in patient setup or beam delivery and anatomical changes are not readily known. A new product, PerFRACTION (Sun Nuclear Corporation), is designed to identify these errors by comparing the exit dose image measured on an electronic portal imaging device (EPID) from each field of each fraction to those from baseline fraction images. This work investigates the sensitivity of PerFRACTION to detect the deviation caused by these errors in a variety of realistic scenarios. Integrated EPID images were acquired in clinical mode and saved in ARIA. PerFRACTION automatically pulled the images into its database and performed the user‐defined comparison. We induced errors of 1 mm and greater in jaw, multileaf collimator (MLC), and couch position, 1° and greater in collimation rotation (patient yaw), 0.5–1.5% in machine output, rail position, and setup errors of 1–2 mm shifts and 0.5–1° roll rotation. The planning techniques included static, intensity modulated radiation therapy (IMRT) and VMAT fields. Rectangular solid water phantom or anthropomorphic head phantom were used in the beam path in the delivery of some fields. PerFRACTION detected position errors of the jaws, MLC, and couch with an accuracy of better than 0.4 mm, and 0.5° for collimator rotation error and detected the machine output error within 0.2%. The rail position error resulted in PerFRACTION detected dose deviations up to 8% and 3% in open field and VMAT field delivery, respectively. PerFRACTION detected induced errors in IMRT fields within 2.2% of the gamma passing rate using an independent conventional analysis. Using an anthropomorphic phantom, setup errors as small as 1 mm and 0.5° were detected. Our work demonstrates that PerFRACTION, using integrated EPID image, is sensitive enough to identify positional, angular, and dosimetric errors.


| INTRODUCTION
Medical physicists perform a wide array of quality assurance (QA) measures in support of all patient treatments as well as those that are patient-specific prior to the start of treatment. However, once treatment has started, other than weekly chart checks, there are few if any efforts to verify ongoing patient-specific treatment delivery accuracy. With the advent of complex treatments and tightening target volume margins, image guided treatments are being performed more frequently with the objective of assuring isocenter positional accuracy and reproducible body pose. These efforts, while necessary, are not sufficient to assure that the correct radiation dose is being delivered daily.
The checks that are done prior to the start of the patient's treatment, such as chart and plan checks, and patient-specific QA will not catch errors caused by patient anatomy changes, patient setup errors, and machine output errors. Patient-specific intensity modulated radiation therapy (IMRT) QA tests involving 2D Gamma passing rates commonly done prior to start of treatment have been found to be insufficient to verify the actual dose received by the patient. 1,2 In vivo dose verification has been performed to verify delivered dose, typically only during the first fraction using point dose detectors such as diode, thermoluminescent dosimeters and optically stimulated luminescent dosimeters, and metal-oxide semiconductor field effect transistor. [3][4][5][6] However, a point dosimeter can easily miss the errors that affect the area outside of the measurement point and can be insensitive to small errors because of placement uncertainty and movement due to patient breathing. It typically requires labor for placement, pretreatment calibration, and posttreatment readout.
It also has dependence on some of treatment parameters such as accumulated dose, energy, SSD, field size, linearity, angular orientation, and readout delay, and in general, a point dosimeter has a measurement uncertain up to 3-5%. 6 The Electronic Portal Image Device (EPID) has the advantage of being integrated into most linear accelerators (linac) and is ready to measure QA plans or patient exit dose during treatment delivery.
With submillimeter spatial resolution, and excellent dose measurement accuracy, linearity to dose and dose rate, and capability of collecting the integrated signal or dynamic signal, the EPID has been widely used for machine QA and pretreatment verification such as patient-specific IMRT verification. [7][8][9][10][11] Recently many authors have investigated using EPID for in vivo dosimetry. 6,[12][13][14] Some authors compared reconstructed EPID-based 3D dose distribution inside the patient to the original treatment plan, 6,13,14 and some authors compared the EPID-measured doses to the predicted doses at the EPID level. 12 In addition, some authors implemented real time dose delivery verification by comparing EPID-measured images to calculated model-generated transit EPID images. 12 Most of these prior efforts have been manually performed.
PerFRACTION (Sun Nuclear Corporation, Melborne, FL, USA) is a system that automatically monitors the consistency of daily treatment delivery using the EPID. PerFRACTION automatically retrieves the EPID exit dose images from the radiotherapy electronic medical record (EMR) system database after each treatment fraction for each patient monitored by the system. A user-defined comparison test such as the gamma analysis is performed for each beam and each fraction against a user-defined baseline fraction. Using the EPID images, PerFRACTION has the potential to identify changes in patient anatomy, patient setup, beam delivery, or couch rail positions that could affect the treatment delivery. 15 In order for the test results to be meaningful, the accuracy and sensitivity of the system to measure the changes in dose that can occur during treatment must be characterized. In this work, we investigate the sensitivity of PerFRACTION to detect the deviation of induced errors in a variety of realistic scenarios. As far as we can tell, this is the first publication which presents such data.

2.A | PerFRACTION system overview
The system (PerFRACTION version 1) consists of a dedicated server running embedded Microsoft windows, database software, and a web interface for configuration and data analysis. A DICOM file transfer connection is made between PerFRACTION and the user EMR in our case, ARIA (Varian Medical Systems, Palo Alta, CA, USA).
An integrated treatment beam image is taken for each field for each treatment. The acquired EPID images are automatically saved in ARIA and PerFRACTION automatically retrieves the images from ARIA using an automated query retrieve process. Images of each field measured by the EPID during the first fraction are normally chosen to be baseline images, and images captured during each subsequent fraction are compared against the baseline images. The process requires minimal effort because PerFRACTION automatically compares new images against baseline images using various userdefined tests including the Gamma analysis 16 with user-defined percent dose difference (DD), distance-to-agreement tolerances, dose threshold, and passing rate that is a percentage of pixels passing the criteria. If the percentage of passing points does not meet a preset passing rate (i.e. 95%), PerFRACTION will notify the physicist via email. A web-based interface can also be used to review results for each field for each fraction. More recent versions of the system can also perform 3D dose calculations using cine images of multileaf col-  Avondale, PA) replacement for the standard Varian couch top, which included the Dosemax couch insert and movable rails. Images were saved in the ARIA database in the same way as any patient treatment images. All images were measured at the source to detector distance (SDD) of 150 cm and were not scaled back to 100 cm SDD except for those in the MLC errors with IMRT test (Section 2.C.8) that were analyzed using PerFRACTION version2. To evaluate PerFRACTION's sensitivity in detecting geometric errors such as jaw position, MLC position, collimator rotation, and couch shift errors as in Section 2.C.1-3, and 2.C.5, we used the gamma analyses in PerFRACTION, but we suppressed the DD aspect by setting the DD tolerance to zero, so the tolerance for distanceto-agreement (DTA) determines if a pixel passes or fails the comparison criteria. We refer to this method as the DTA method in this paper, which is defined as the distance between a point in an image compared to the nearest point with the same dose in another image.

2.C | Experimental design
If the DTA value is out of tolerance, the gamma value is >1; if the distance is within tolerance, the gamma value is ≤1. PerFRACTION renders the pixel to be orange if the gamma is >1 and renders a pixel to be between green and yellow for gamma between 0 and 1.
We first set the DTA tolerance to be greater than the value of an induced error for which we expected no failing pixels. Then, we decreased the DTA tolerance in 0.1 mm increments until failing pixels began to appear in the color map, and recorded the final tolerance value. In all cases, failing pixels only occurred with DTA tolerances less than the induced error. The difference between an actual induced error and the tolerance implies the sensitivity of the       The gantry was at 45°(IEC) and 100 MUs was delivered with each field. A PerFRACTION-calculated DD method was used for error analyses.

2.C.3 | Collimator rotation errors
2.C.7 | Rail position error with a VMAT field A 10 cm 9 10 cm VMAT field was delivered with the arc passing through the couch and the rails at the outer-most position. The isocenter was at the couch surface at its center. Then the field was delivered again with the rails moved to the inner-most position of the couch. 300 MUs was delivered with each arc field. A PerFRAC-TION-calculated DD method was used for error analyses. that with induced errors. These dose grids were calculated from a 3D dose grid with a 2 mm resolution, resampled to 0.39 mm resolution and exported to the SNC Patient software to perform the gamma analyses using 2%/3 mm since the calculation plane was 150 cm from the source and not scaled back to 100 cm. Thus, the effective depth of measurement for all three analyses was close to d max , the beams passed through the same 20 cm thick solid water slab, the source and phantom to detector or calculation plane was the same, the gamma analysis parameters were the same for all images when scaled back to 100 cm distance, and the dose grid resolutions were nearly the same. The gamma passing rates for each field for the Eclipse calculations, MapCHECK2 measurements, and the PerFRACTION measurements were compared.

3.A | Jaw position errors
The DTA tolerance was decreased as described in the Methods section until failing pixels were seen. The smallest jaw position shift, which is 1.5 mm in the EPID image, was apparent with the DTA tolerance set to 1.3 mm. Fig. 1

3.B | MLC position error
The smallest leaf position error, which is 1.5 mm in the EPID image, became apparent, appearing yellow or orange in the color map when the DTA tolerance was set to 1.1 mm. Fig. 2 shows that the Per-

3.D | Machine output error
For induced errors of 0.5%, 1.0%, and 1.5%, PerFRACTION DD showed failing pixels when the DD tolerance was set to 0.5%, 1.2%, and 1.6%, respectively. The sensitivity of PerFRACTION to identify an output error is 0.2%.

3.E | Couch position errors
For the induced couch position shift, which was at a minimum 1.5 mm in the EPID image, PerFRACTION began to display failing pixels when the DTA tolerance was set to 1.7 mm. The sensitivity of PerFRACTION in identifying a couch position error is therefore 0.2 mm. Fig. 4(a,b) shows the changes in the integrated image due to rail position error with a static field. When the DD tolerance is set to any value up to 8%, PerFRACTION renders blue and red pixels in the color map (Fig. 4c,d), meaning the maximum DD is about 8%.

3.F | Rail position error with an open field
This test indicates that PerFRACTION can identify a rail position error with an open static field irradiation, and in this case represented a dose error up to 8%.
3.G | Rail position error with a VMAT field Fig. 5(a,b) display the integrated exit dose images with rails in or out with a VMAT field irradiation. The difference between Fig. 5(a,b) is not visually obvious but can be quantitated by PerFRACTION. As shown in Fig. 5(c-f), PerFRACTION displays a few blue or red pixels with DD tolerance set to 3%, and displays more failing pixels with reduced DD tolerances. This test indicates that PerFRACTION is sensitive in identifying a rail position error during VMAT delivery, with as much as a 3% dose error being uncovered in this case.

3.J | Constancy check
The constancy check showed that the EPID-measured integrated exit dose images were consistent with a deviation of 0.2% or less in dose. With this small error in reproducibility of the linac output, we made no corrections for this effect.

| DISCUSSION
Interpretation of exit dosimetry results depends on the sensitivity of the system to detect an error. Ideally, the detection system would be able to discern errors much smaller than are clinically rel-  We induced very small to moderate lateral and longitudinal shifts and rotational errors during the setup of a head phantom to mimic realistic clinical situations. Because the anatomy of the head phantom contains heterogeneous tissues such as soft tissue, air, and bone, when the head phantom was shifted from baseline position, the pattern of the exit dose image changed. PerFRACTION gamma analysis was able to demonstrate even the smallest induced errors if correspondingly tight tolerance levels were used in the analysis, demonstrating that PerFRACTION has the sensitivity to be able to alert the user to errors that are even smaller than might be considered clinically significant. Actual patient results using PerFRACTION will be reported in a separate publication.
Acquiring EPID images for PerFRACTION is limited to couchgantry angle combinations that don't cause imager to couch or patient collisions. For coplanar beams, 150 cm source-imager distance allows imaging even for most off center couch positions. For noncoplanar beams, increasing the source-imager distance from 150 to 170 cm greatly increases the range of beams that can be imaged, but there will still be those that could cause a collision. In our experience, for a typical 10 beam noncoplanar plan, at most two beams cannot be safely imaged.

| CONCLUSION S
The PerFRACTION system, which is comprised of software that automatically retrieves EPID images for each fraction and compares them to the baseline, typically the first fraction, is sensitive enough to provide useful and actionable information about the reproducibility of treatment delivery and patient setup. This type of fully automated daily patient treatment QA using the ubiquitous EPID, is feasible since it uses virtually no physicist's time and fills an important unmet need for a better understanding of the accuracy of daily treatment.

CONFLI CT OF INTEREST
Dr. Arthur Olch receives research funding from Sun Nuclear Corporation but no funding for this work was provided.