EPID-based in vivo dosimetry – new developments and applications

In vivo dosimetry has been shown to be a powerful quality assurance method in modern radiation therapy. The most common tool used for in vivo dosimetry is the electronic portal imaging device (EPID) which can quantitatively image the therapeutic beam fluence exiting the patient during treatment delivery. Since the last major literature review on this topic was published five years ago, the radiation oncology community has shown continued strong interest in this subject. Commercial options have become more widely available, with a related increase in validation efforts and sensitivity testing, while new applications continue to be explored. Work has been done to understand and increase the accuracy of the EPID for dosimetric applications, as well as continued efforts to provide practical, quantitative experiences from clinical implementation of in vivo dosimetry systems. This review examines the published literature related to in vivo EPID dosimetry from January 2017 to February 2022. The literature is classified into three main topical areas: (1) new or improved algorithmic developments including validation work, (2) applications of the in vivo EPID dosimetry method, and (3) error identification and error sensitivity analyses.


Introduction
Modern radiation therapy continues to witness a rapid pace of technological development which allows advances in clinical procedures delivering higher dose in fewer fractions and to increasing levels of accuracy and precision.However, differences between the intended delivery and the actual delivery can still occur.In vivo dosimetry (IVD) is a valuable tool that allows a measurement verification of the actual dose delivered to the patient.A recent review paper by an ESTRO task group of experts provided a standard definition of in vivo dosimetry: "IVD is a radiation measurement that is acquired while the patient is being treated containing information related to the absorbed dose in the patient."[1].Note that this definition excludes related and potentially useful but often misperceived techniques such as monitoring the constancy of transmission images as well as log-file based analysis.The in vivo dosimetry approach is able to detect errors in machine/hardware delivery, errors in dose calculation, errors in patient setup, and changes in patient anatomy between planning and delivery.It has been demonstrated to be one of the most important quality control checks in radiation therapy [2], and is recommended by several organizations to increase patient safety (e.g.WHO [3], ESTRO [1], and AAPM [4]).
In the 1980's and 1990's, there was interest in using electronic portal imaging devices (EPIDs) for in vivo dosimetry measurements.When the current generation amorphous-silicon style EPIDs became widely available in the 2000's, interest grew dramatically due to their excellent dosimetric properties [5,6] which were significantly improved over previous EPID technologies (both camerabased and liquid ionization chamber designs).The EPID's broad availability on modern linear accelerators, convenience of use, large 2D detector format, and digital image nature ensure that EPIDbased in vivo dosimetry is significantly more attractive compared to alternative, manually placed, single point detector methods.
Several excellent reviews of a variety of aspects of EPID based in vivo dosimetry are available in the literature.Van Elmpt et al provided a comprehensive literature review in 2008 [7], which was updated to late 2016 [8].In 2020, an ESTRO task group on in vivo dosimetry provided a broad assessment of the current state of in vivo dosimetry clinical implementation, as well as describing areas requiring more development, and suggested approaches to improve clinical acceptance [1].There have also been recent reviews of EPID-based in vivo dosimetry focusing on SBRT delivery techniques including for lung treatment, where available techniques and clinical dosimetry results were summarized in 2017 [9], and for all SBRT treatment sites, providing an overview of accuracy and sensitivity of published results in 2020 [10].Other related reviews include a report on clinical usage of in vivo dosimetry methods in the UK [11] as well as in vivo dosimetry methods in brachytherapy [12].
The purpose of the present review of EPID based in vivo dosimetry is to update the literature reviews of references [1] and [8] to the period of January 2017 to February 2022, focusing on in vivo dosimetry.Three main categories will be examined: (1) new or improved algorithmic developments including validation work, (2) new applications of the IVD method, and (3) error identification and error sensitivity analyses.It is recognized that some papers qualify in multiple categories, in which case a decision was made as to the best single category placement.

Predictive methods or algorithms
Several groups have contributed additional empirical methods aimed at estimating dose at a 2D plane within the patient, generally located at mid-plane.An existing empirical method to estimate 2D midplane dose in the patient was improved through use of a more accurate off-axis correction method, incorporating an asymmetric 2D Gaussian kernel to account for beam penumbra effects [13], with the MatLab code (MathWorks, Natick, CT) made available to researchers.Hashemi et al [14] developed an empirical method for homogeneous phantoms accounting for beam hardening, patient scatter, and off-axis EPID response, and verified it using a 6MV, 8x8 cm 2 field on a 20 cm thick water equivalent slab comparing reconstruction to a treatment planning system (TPS) with a gamma pass rate (3%, 3 mm DTA) of 94.9% for points at mid-plane.This method was then extended with a water-equivalent pathlength concept to account for patient CT data [15] and tested on a pelvis phantom with a conformal field and a 4-segment step-and-shoot IMRT field, with gamma results (3%, 3 mm DTA) within 97.0% and 93.5% respectively.Anvari et al [16] proposed an empirical method with attenuation correction and off-axis correction terms to estimate 2D mid-plane dose and also EPID plane dose and tested it with a homogeneous phantom using a fluoroscopic style EPID as well as a planar dose array, observing a gamma pass rate (3%, 3 mm DTA) of 96% for fields from pelvic plans.
A 3D backprojection method originally developed by Berry et al [17] was improved by the addition of a correction factor to remove dependence on source-imager-distance, and by incorporating a model of the treatment couch [18].The method was tested with six IMRT plans and seven VMAT plans on a homogenous slab phantom, RT01 pelvis phantom, and anthropomorphic pelvis phantom, with gamma pass rates (3%, 3 mm DTA) of >95% for all the IMRT fields and >95% for 23 of 24 VMAT arcs.
Results for a 2D forward projection approach using a Monte Carlo method have been recently published by Feng et al [19] using a fast simulation method based on Jia et al [20], with average gamma pass rates (3%, 3 mm DTA) for 268 fractions over 20 IMRT patients of 87.8% (±14.0%) using the planning CT data set and 93.6% (±8.2%) using the daily CBCT data set.
There is continued interest in utilizing commercial treatment planning systems (TPS) to calculate dose to the EPID plane.Deshpande et al [21] used Pinnacle (v9.1) and introduced a waterequivalent virtual EPID into the TPS and compared predictions to measurements using a commercial EPID that they physically converted to a water-equivalent EPID by replacing the metal/phosphor buildup layers with solid water.Gamma pass rates (3%, 3 mm DTA) of >94% were achieved for IMRT and VMAT fields applied to inhomogeneous and anthropomorphic phantoms.Martinez-Ortega et al [22] used Pinnacle (v8.0m) and introduced a water-equivalent virtual EPID model to the TPS but also implemented a photon spectral correction based on radiological thickness of the patient along individual raylines.They compared predictions to a-Si EPID measurements that were converted to be water-equivalent, and achieved gamma pass rates (5%, 3 mm DTA) of >80.3% for three square fields on an anthropomorphic phantom (cranial, thorax, and pelvis sites) and >94.6% for five IMRT fields on the pelvis site.Antoniou and Penfold developed a script to add a virtual EPID model consisting of a 5cm thick water slab backed by a 4 cm thick lead slab, into the patient calculation volume using Raystation (v8B) [23].Mean differences with a measured integrated EPID image for a VMAT delivery were within 9.0%, with the mean difference across the entire image of 1.5%, although the investigation was more focused on establishing correlations between transit fluence changes and DVH metrics.
Guo et al developed a new hybrid method for estimating the patient-generated photon scatter fluence entering the EPID [24] combining three different algorithms suited for different photon scatter types: analytical for singly scattered photons, hybrid Monte Carlo for multiply scattered photons, and pencil beam for bremsstrahlung and positron annihilation photons.The authors also optimized it's sampling for speed and accuracy, achieving the accuracy of full Monte Carlo simulation at a fraction of the time [25].That algorithm can be used as part of the CancerCare Manitoba (CCMB) inhouse 3D dose reconstruction technique to improve overall accuracy.The Netherlands Cancer Institute (NKI) group developed a more accurate version of their inhouse algorithm specific to small field dosimetry (ie.stereotactic body radiotherapy or SBRT) by removing the influence of the ion chamber volume effect from the EPID scatter kernel, extending output factors to smaller fields, and updating scatter-toprimary ratios using smaller incident fields [26], changes which reduced dose discrepancies in tested brain VMAT fields from 6.8% to 0.8% at isocentre.They also extended their in vivo EPID dosimetry technique to an MR-linac (Elekta Unity, Elekta AB, Sweden), documented in several publications as follows.In 2017, Torres-Xirau et al demonstrated the feasibility of modifying the EPID dose prediction method to account for attenuation and scattering of the cryostat in an MR-linac, simulated with a conventional linac [27].Once the MR-linac was available for testing, the authors provided a more detailed update of the NKI in vivo EPID model [28], specifically for EPID plane dose prediction.Twenty-five IMRT fields (two prostate plans, one rectum plan) were tested with gamma test results (3%, 3 mm DTA) all >95%.This was eventually extended to 3D patient dose reconstruction [29] with plans from five patients being tested comparing EPID reconstructed dose with in-phantom measured dose (Octavius, PTW, Frieburg, Germany), achieving a gamma pass rate (3%, 2 mm DTA) average of 91.9%.However, to improve agreement further especially for larger fields whose outer regions are more significantly affected by attenuation and scatter from gradient coils, Olaciregui-Ruiz et al [30] used a deep learning approach to correct in vivo reconstructed dose in these areas, demonstrating the average gamma pass rate (3%, 2 mm DTA) outside of the central region improved from 44.5% to 90.2%, using 45 large field rectum plans.

Validation of algorithms
The last five years has seen an increase in the interest in and use of commercial options for in vivo dosimetry, as demonstrated by numerous validation studies available in the literature during this time.Stevens et al [31] validated Dosimetry Check (Lifeline Software, Tyler, TX), a 3D patient dose reconstruction software, using PAGAT gel dosimetry in two phantoms using a single VMAT plan, and obtained gamma pass rates (3%, 3 mm DTA) of >92% and >94%.Esposito et al [32] validated Dosimetry Check with Octavius phantom measurements for 20 clinical VMAT plans, demonstrating a mean gamma pass rate (3%, 3 mm DTA) of 94.2% (±3.4%).They also evaluated error sensitivity to various setup errors and anatomical changes.Held et al [33] validated EpiGray (DOSIsoft, Paris, France) using solid water phantoms of three thicknesses and a variety of field shapes including IMRT.Dose reconstructed at a single point (typically isocentre) was found to be within 3-5% of the TPS, while off-isocentre points typically showed poorer agreement.The PerFraction software (Sun Nuclear, Melbourne, FL) was validated by Aziz-Sait et al [34] using an ion chamber and a 2D ion chamber array (MapCheck, Sun Nuclear) to validate single arc VMAT plans in several phantoms (homogeneous and inhomogeneous) and found 97.8% of 270 compared points were within 3% of expected.Olaciregui-Ruiz et al [35] validated the iViewDose software (Elekta AB, Stockholm, Sweden) with Octavius phantom measurements using 10 IMRT plans and 34 VMAT plans (2 arcs each), finding an average gamma pass rate (3%, 2 mm DTA) of 92.2% (±5.2%).
The CCMB inhouse method was validated for in vivo dose reconstruction on 4D CT data sets using a dynamic thorax phantom and lung tumour insert [36] showing a gamma pass rate (3%, 3 mm DTA) within the PTV of >99.3%.The method was also validated for non-coplanar beam configurations such as those used in stereotactic radiosurgery (SRS) brain treatments [37], using an IMRT QA phantom with irradiations delivered via XML scripting in research mode on a TrueBeam linac (Varian Medical Systems, Palo Alto, CA), achieving gamma pass rates (2%, 2 mm DTA) >97%.

Infrastructure work
This section describes development work that is not itself a type of in vivo dose calculation, but rather relates to required methodological infrastructure or foundational understanding which impacts the accuracy of an implemented in vivo dosimetry method.
For example, the pixel sensitivity matrix (PSM) is a technique used to correct for the variation in gain of individual pixels throughout the EPID panel, while allowing the image to still retain beam profile dosimetric information.Accounting for this effect improves the accuracy of EPID in vivo dosimetry applications by up to ~5%.Cai et al proposed an approach acquiring five, large-field overlapping irradiations where the EPID is shifted by a few mm between irradiations [38].Beam on images were alternately acquired with dark field images, which were later subtracted in order to eliminate ghosting effects.Another proposed approach avoided shifting the EPID and used a regression model to differentiate the shape of beam fluence (i.e.low frequency signal) from the pixelto-pixel variation (i.e.high frequency signal) [39], thus deriving the PSM in a single EPID image acquisition.
Abbasian et al [40] showed that one could measure the timing and amplitude of linac dose pulses and use this to modify a 2D EPID dose image prediction method to account for the spatialtemporal readout nature of the Varian EPID.Improved 2D image prediction for individual EPID frames was demonstrated, of interest for real-time image comparison applications.
Renaud et al [41] performed a systematic evaluation of the dosimetric uncertainty for a unique film-in-EPID dose calibration method for EPIDs (Elekta model examined).
The same authors also evaluated long term stability of the Elekta EPID panel and showed it to be dependent on supply voltage, internal operating temperature, and accumulated absorbed dose in the EPID [42].Barnes et al [43] demonstrated the long-term drift of sensitivity using a Varian EPID and recommended regular (i.e.monthly) cross calibration with ion chamber measurement.

2D forward dose projection in the EPID
Fuangrod et al [44] applied statistical process control methods to the comparison of measured and predicted transmission EPID images for IMRT and VMAT treatments to flag treatment discrepancies which required investigation.Control limits were generated using initial data from 90 IMRT and 10 VMAT patient deliveries.Fourteen patients were selected as case studies, and five patients had treatments flagged when verification exceeded the control limits.Investigation of the flagged treatments revealed that most errors were due to anatomical uncertainty.

Dose reconstruction in the patient for external beam treatments
There has been significant effort in publishing clinical experiences with in vivo EPID dosimetry programs over the literature review period.Note that many of these are using commercial products while only a few use inhouse developed techniques, a change from earlier literature reviews, which had very few results from commercial products.
The Softdiso system (Best Nomos, Pittsburgh, PA, USA) has been applied to a group of 15 lung SBRT patients (80 fxs, 2 arcs/fx) [45], and observed an out-of-tolerance rate of 5%.The system has also been used on a mix of 3DCRT and VMAT deliveries across breast, thorax, abdomen, pelvis, H&N, and brain (823 patients, 11357 fxs) [46], finding 21% of 3DCRT fractions and 6% of VMAT fractions out-of-tolerance, with the higher value for 3DCRT attributed mainly to less consistent patient setup and also attenuation through uncontrolled positioning of couch supports.The Softdiso system was also applied to a mix of 3DCRT (breast, lung, pelvis, abdomen), IMRT (pelvis, H&N), and VMAT deliveries (prostate, H&N) on a total of 386 pts, 8474 fxs, and observed 9%, 7%, and 9% outof-tolerance rates respectively [47], with the results of initial fractions used to identify and correct errors in subsequent fractions to bring them within tolerance.
Nailon et al [48] applied Dosimetry Check to verify in vivo dose at the reference point for 2943 3DCRT and 842 VMAT patients and found an out-of-tolerance rate of 4%.Esposito et al [49] evaluated Dosimetry Check on abdominal and prostate SBRT patients (80 pts, 152 fxs), observing an out-of-tolerance rate of 7%.A pre-cursor to the iViewDose software, modified to work on the Unity MR-linac system (Elekta), was applied to 127 IMRT treatments (74 prostate, 19 rectum, 19 liver, and 15 lymph node oligomets) during 1207 adapted treatment fractions, with a 9% out-of-tolerance rate observed [50].Bossuyut et al [51] used the PerFraction software for in vivo dosimetry (and pretreatment QA) and reported in vivo results from 24011 fractions across 3671 patients (used on first three days then weekly thereafter) over a wide range of anatomical treatment sites using integrated EPID transmission images, as well as discussing error types and the development of clinical tolerances.For their in vivo dosimetry results they found a 16% out-of-tolerance rate, of which 6% were false positives and 10% were related to patient setup, with newer linacs associated with about half the false positive rate of older linacs.
Clinical results from non-commercial algorithms include McCowan et al [52] using the CCMB inhouse method on 602 fractions delivered to 117 lung, spine, and liver SBRT patients, and demonstrating an out-of-tolerance rate of 10% after EPID frame averaging was optimized.Meanwhile Peca et al [53] used their 2D midplane algorithm on 10 rectal patients treated with 4-field box plans and observed out-of-tolerance differences in seven patients due to issues with positioning of the bellyboard, patient positioning on the belly-board, and bowel gas.Sterkx et al [54] used the 'in aqua' in vivo EPID dosimetry method [55], originally developed for lung patients, on prostate patients who were treated with a rectal balloon device to circumvent the negative impact of the presence of air or low density anatomical volumes to accurate patient dose reconstruction.Anvari et al [56] used their inhouse method, an empirical 2D mid-plane dose reconstruction [16], to verify small animal (ie.rat) irradiations showing agreement of TPS predicted and EPID measured central axis exit dose rates of 3.1% (on average).The literature review also showed that many researchers are using available in vivo EPID dosimetry systems as a tool to investigate other aspects of radiation therapy, beyond the evaluation or quantification of the in vivo dosimetry system itself.For example, the Softdiso system was used to demonstrate and quantify improvements in lung SBRT treatment reproducibility in 10 patients when using the ABC breathing system [57], as well as to assess the treatment quality in 48 breast patients and correlate this to an improved patient setup procedure [58].The EpiGray software was used to help quantitatively evaluate the dosimetric impact of high-density bone cement [59] during radiation therapy, of interest for post-surgical spine SBRT patients.The PerFraction software was applied to quantitatively evaluate the effectiveness of new patient preparation strategies for the bladder and rectum in prostate SBRT patients, and while was shown to detect changes in bladder and rectal filling, could not confirm the hypothesis [60].Tan et al [61] used the non-commercial in vivo EPID dosimetry approach of Najem et al [18] to help develop a new quality assurance methodology to determine the target localization accuracy of phase gated treatments.

Dose reconstruction in the patient for brachytherapy treatments
Smith et al [62] verified dwell times and source positions for HDR treatments of two prostate patients using an external imaging panel (aS500, Varian) embedded in the treatment couchtop using the HDR source as an x-ray imaging source.They used a pre-treatment localization planar x-ray image to establish prostate position (using fiducials) and catheter positions (using x-ray dwell position markers).In a proof-of-concept study, Fonseca et al [63] also used an external EPID and combined the HDR source tracking with anatomical gamma ray imaging by the source itself, to remove the need for pre-treatment planar x-ray imaging, to determine source dwell times and positions for HDR treatments in a pelvic phantom.

Error identification and error sensitivity analyses
Some error sensitivity analyses are included in the initial validation work, but many publications solely focus on evaluating errors and developing a clinically relevant understanding of the types and magnitudes of errors detectable by in vivo EPID dosimetry.
Bedford et al [64] used the iViewDose software and introduced errors into thirteen clinical prostate VMAT plans and delivered these to a phantom in order to compare the results of backprojection in vivo dosimetry with forward projection in vivo dosimetry, demonstrating that while the two methods have their own error sensitivities, they are generally similar.That research group also investigated the use of a running sum of real-time EPID transmission images for comparison to predicted EPID transmission images for corresponding segments of a VMAT arc [65], in a forward projection framework.They introduced errors in MU's, field size, field position, and patient thickness for 13 clinical prostate plans and delivered to a phantom to determine error sensitivity.More recently they developed a composite difference metric to identify forward projection errors more quickly, combining central axis signal, mean image signal, and two different image difference metrics [66].
A pre-cursor to the iViewDose software was used by Van Der Bijl [67] to compare DVH based tolerances to multi-parametric gamma-based tolerances for pelvic VMAT patients (47 rectum, 37 prostate, 44 bladder pts; 387 fx total) and establish meaningful correlations with the goal of establishing more clinically intuitive tolerances.Also using a pre-cursor to the iViewDose software, Mijnheer et al [68] tested sensitivity to a variety of errors (ie.delivery, thickness, setup) physically introduced to treatments on an anthropomorphic phantom using four VMAT plans (prostate, lung, two head and neck).In addition to establishing more clinically relevant tolerance critieria, they also observed that the presence of patient setup errors reduced sensitivity to other types of errors.Again employing similar IVD software, Olaciregui-Ruiz et al [69] mapped the sensitivity for a variety of patient related errors using a virtual phantom approach, where errors are introduced into the CT data set or the treatment plan, and the dosimetry recalculated and compared to the non-errored dosimetry.They developed alert criteria specific to four anatomical treatment sites (using 104 VMAT plans from brain, rectum, head and neck, and lung) and the magnitude of errors they wanted to detect clinically.The virtual phantom approach was also used by that research group to compare error detectability for in vivo dosimetry estimated on the patient's planning CT versus the patient's daily CBCT [70], demonstrating that the use of the daily CBCT was more sensitive in detecting patient positioning errors, which is logical considering it is a more accurate representation of the dose actually delivered to the patient.Yedekci et al [71] evaluated the sensitivity of the iViewDose software on a variety of errors (dose calibration, setup, collimator, MLC, and patient anatomy) introduced to ten prostate SBRT patient plans delivered to a pelvis phantom and showed that some positional errors could escape detection by in vivo dosimetry in that treatment site.Recently Mans et al [72] presented a procedure to perform long-term trend analysis of multi-year in vivo EPID results for prostate cancer patients to identify gradual changes over time, which led to the identification and reduction of underlying systematic dose uncertainties through implementation of new beam fits in the treatment planning system.Esposito et al [73] performed an eight-centre study comparing clinical results (i.e.out-oftolerance rates, errors identified) of three different commercial in vivo dosimetry systems (Softdiso, PerFraction, and Dosimetry Check), involving 32,276 fx and 2002 patients.They showed that the proportions of treatments flagged as out-of-tolerance, and the identified types of errors, were generally consistent between the three software systems.
Wolfs et al [74] used an inhouse, forward projecting EPID algorithm to generate integrated VMAT delivery images with 47 lung cancer patient data sets (59 tx plans and 175 beams).Anatomical, positional, and mechanical errors were introduced virtually via modifications of the patient CT data set and also the treatment plan.A convolutional neural network was trained to identify these errors and showed promising results (96%, 87%, and 73% classification accuracy for increasingly complex errors).While their inhouse in vivo EPID dosimetry method was not described in full detail, Li et al [75] performed a sensitivity assessment on it by introducing several errors (machine output, PTV deformation, MLC, and patient setup) to treatment delivery of one IMRT and one VMAT plan and evaluated on two simple geometric phantoms and one thoracic phantom.

Summary
A literature review focusing on recent EPID-based in vivo dosimetry is presented.'Recent' is defined here as the approximate five-year period from January 2017 to February 2022, considering the last major comprehensive EPID dosimetry review [8] included publications to the end of 2016.The published literature was categorized into three main areas including: (1) new or improved algorithmic developments including validation work, (2) new applications of the IVD method, and (3) error identification and error sensitivity analyses.The number of publications annually on this topic indicates it maintains a strong interest from both the research and clinical communities in radiation oncology.

12th
International Conference on 3D and Advanced Dosimetry Journal of Physics: Conference Series 2630 (2023) 012009 Validation studies typically focus on examining in vivo dosimetry results measured under carefully controlled situations (i.e.phantom irradiation) and comparing those to patient/phantom dosimetry calculated in a treatment planning system.