An electronic portal imaging device (EPID) is routinely fitted as a radiotherapy linear accelerator accessory. The EPID can be used to verify the multi-leaf collimator (MLC) leaf positions and other quality assurance (QA) type tasks. EPID-based MLC has been extensively studied for position error detection of large fields and using a variety of leaf verification algorithms [1–7]. But there are currently no widely accepted EPID methods that can accurately measure position errors for exceptionally small irregular-shaped MLC fields.
New, high precision techniques such as stereotactic radiotherapy (SRT) and stereotactic radiosurgery (SRS), use very small MLC defined fields. For the stereotactic technique, the mult-leaf collimators produce a highly conformal X-ray dose distribution over a very small localized target volume [8] of up to 3 cm. The technique demands very high spatial multi-leaf positional accuracy.
Joshua Hiatt BSc MSc (Medical Physics)
Medical Physics Registrar (Radiation Oncology)
This article is a technical overview of the Better Healthcare Technology 2019 prize-winning journal article:
MLC Positioning Verification for Small Fields: A New Investigation into Automatic EPID-based Verification Methods Joshua Hiatt, Godfrey Mukwada, Michael Barnes, Hans Lynggaard Riis, Du Huynh, Pejman Rowshanfarzad.
The work described in this paper was undertaken as part of MSc research studies at the University of Western Australia.
Introduction
On-going advances in radiotherapy technology has enabled:
- more reproducible and reliable delivery systems;
- smaller EPID pixel sizes for patient or X-ray beam check images;
- improved linear accelerator collimation control systems; and
- the miniaturisation of MLCs.
The smaller MLC leaf widths also enable better conformity around the planning target volume [9]. The stereotactic techniques are also reduced to much less than the conventional number of fractions and the dose administered (typically 6–30 Gy over 1–5 fractions) is intensified. Consequently, stereotactic techniques require more stringent QA procedures and tighter mechanical tolerances than conventional radiotherapy [10, 11].
And, as radiotherapy X-ray techniques continue to become increasingly complicated, physics QA to check the accuracy and monitor the reliability of the radiation dose, can become more complex and time-consuming.
The objective of this study was to assess available algorithms used in other fields of computer imaging programs that could be used to detect sub-pixel field edges recorded in EPID images. The eventual aim is to develop an automatic, independent EPID-based method that can accurately verify the MLC leaf positions in near real-time during stereotactic (SRS/SRT) treatment exposures.
The potential negative dosimetric impact of positioning errors in small stereotactic fields is relatively large. In fact, for field sizes, less than 12 mm, a 1 mm error in MLC positioning can result in an error in the central axis dose to be greater than 1.7% [12, 13].
According to AAPM TG-142, the positioning accuracy of MLC leaves should be within ± 1 mm [1]. Given the sensitive nature of SRT, it may be prudent to adopt the stricter tolerance in leaf positioning suggested by ESTRO of ± 0.5 mm [10, 14].
Early linear accelerator collimators
Early linear accelerator collimators
The MLC was invented in Japan by Takahashi in 1965 [15], but was not widely implemented until the mid-’80s. As such, X-ray beam defining tests for less sophisticated collimation systems were well established prior to the MLC introduction. Regular shaped rectangular or square X-ray beam sizes were controlled by the jaw collimator system. The X-ray beam could only be further ‘shaped’ by physically attaching lead blocks or shield cut-outs in the path of the beam.
Definition of Collimator Field Size
The X-ray field size for a linear accelerator is defined by measuring the full-width at half-maximum (FWHM) of the X-ray beam dose profile in a water phantom (Figure 1.).
The positional accuracy, field size calibration and symmetrical alignment of the machine’s main collimators are extensively tested during the linear accelerator’s initial detailed acceptance & commissioning program. Additionally, the axis of rotation of the mechanical, optical and X-ray beam are all thoroughly tested for coincident alignment. On-going collimator QA tests can be regularly checked by taking a film exposure of the X-ray beam and examining the beam edge measurements rotating about the collimator central axis.
X-ray Beam Shaping using the Multi-leaf Collimator
Figure 3 shows a Varian linear accelerator MLC system. The leaves are positioned to produce a ‘figure 8’ X-ray beam test field. The jaw collimators (visible above and below) define the outer perimeter of the X-ray beam shaped by the MLC;
The linacs have their own in-built system to monitor the positioning of the MLCs. Varian systems utilise a driving step motor to position each MLC. The number of motor counts to position each leaf is calibrated by pushing each leaf into the path of an in-built infra-red beam and photodetector system. A secondary feedback system monitors the voltage across potentiometers – the resistance of which changes as a function of the MLC position [9]. Elekta systems utilise a curtain of light in the UV spectrum, that is fluoresced into infra-red light from ruby reflectors attached to the MLCs. The fluoresced light is subsequently imaged by a camera [16]. Irrespective of the technology used by the manufacturer to calibrate and position the collimation, it is important to independently verify and monitor the accuracy of these systems.
The MLC leaves define the beam size and shape by driving the leaf pairs individually to a calculated co-ordinate position in the X-ray beam. Each leaf can be programmed to move in increments of 0.1 mm and has a projected width of 0.5 mm at the iso-plane. The X-ray field, in this example, is centred about the central axis.
Electronic Portal Image Device (EPID) for MLC Quality Assurance
Similar to how the EPID can be used for checking the patient’s treatment set-up, the EPID can be used for MLC leaf position QA checks. A number of algorithms have been implemented to verify the position of the leaf collimators, including the use of FWHM measured via EPID pixel intensity. As well as edge detection using methods such as image derivative extrema, template matching, etc…. (see below).
The co-ordinates of the MLC, recorded on the EPID image, can provide a convenient, efficient, and (importantly) independent method for checking the calibration settings of each collimator leaf pair. Although attached to the linac, the effect of gravitational forces, panel flex, and mechanical wear mean that the pixel position of each collimator component needs to be related to a reference point in space because the EPID panel may move relative to the linac’s 3-D coordinate system. In this study, the radiation field centre (at each gantry angle and determined from two collimator angles opposed 180° apart) served this function.
Automated QA methods are now even more important because modern radiotherapy has developed a wide variety of dynamically delivered small ‘pencil beam size’ X-ray treatment fields – referred to as ‘Intensity Modulated Radiotherapy’ (IMRT).
The very small pencil beams are either (i) continually moving (sliding window) while varying the size and shape of these small beams or (ii) deliver multiple precise small beam shapes (step and shoot) across the total treatment field. Small positional errors of the leaves can cause noticeable errors in the dose delivered across the treatment field, especially if there is a systematic component to them.
Special requirements for stereotactic treatments
The use of the EPID to detect leaf positional errors in relatively large X-ray beam fields was extensively studied and published by a number of researchers. References (1-7) list a variety of leaf verification algorithms. However, there are currently no routinely used EPID methods with a sufficiently high degree of accuracy (that can record in real-time) a measure of MLC positional errors in small, irregularly shaped X-ray fields.
Stereotactic Radiotherapy (SRT) and Stereotactic Radiosurgery (SRS) require even higher precision again than intensity-modulated MLC techniques. For a more detailed description of Stereotactic Radiotherapy (SRS) and Stereotactic Ablative radiotherapy (SABR) techniques, see previously published better healthcare technology articles:
What is Stereotactic Radiotherapy?
and:
SABR: A new force in local ablative cancer treatment?
It’s important to reiterate the special requirements of stereotactic treatments. The technique:
(i) uses a highly complex, conformal X-ray dose distribution to a very small localized target volume [8]; and
(ii) is prescribed to give an escalated dose in fewer fractions than is conventionally given.
To successfully achieve this, the MLC leaves must be very accurately driven dynamically to the correct position when changing the shape of these small fields during the total X-ray beam exposure.
Limitations of small field dosimetry
A 6 MV photon of less than 3 × 3 cm2 field size is considered to be a ‘small’ field [18]. For fields of less than this size, phenoma such as:
- occlusion of the detector from the X-ray radiation source; and/or
- lateral charged particle dis-equilibrium;
occur.
Under these conditions, the traditional FWHM measured by the EPID detector will no longer be an accurate measure of the MLC defined field size. Source occlusion, electronic dis-equilibrium, and the penumbra overlap on the EPID created by each opposing MLC leaf create a distorted result. The combined effects appear to cause an apparent widening of the field’s FWHM (18,19).
There needs to be:
- A better understanding of the physics dosimetry;
- A more proficient method to accurately measure the MLC leaf gap;
- Much more stringent QA tests with tighter tolerances; and
- A means of accurately verifying MLC leaf position.
Despite limitations in using the EPID to accurately quantify dose for small fields, EPID scans can be used to accurately measure the position for a given MLC leaf number.
The cross-profile X-ray intensity (measured by the EPID matrix detectors) provides a convenient means of measuring the x,y pixel coordinate position for each leaf pair. The computer software is programmed so that the radiation field centre (0,0,0) determined on the EPID at each gantry angle can be related to the leaf position (x) and the leaf number (y).
The MLC field size = the separation between the FWHM EPID intensities of opposing Leaves.
The Y position is determined from the field height (number of leaves used to define the field).
EPID Co-ordinates versus IEC Watertank Coordinates
The EPID intensity profile measurement is technically not the same as the IEC water-tank co-ordinates. Also, the EPID ‘pixel intensities’ are not readily convertible into a measure of dose. EPID field size should be more correctly referred to as the distance between the 50% iso-intensity points, rather than 50 % isodose.
Dose errors for Small Field Leaf Errors
Dose errors arising from MLC leaf errors are far more significant in high dose, very small field size, stereotactic techniques. For example:
- a 0.5 mm error in the solid collimator setting is considerably larger for a 1 cm2 field size than for a 10 x 10 cm2field; and
- for field sizes ≤ 12 mm, a 1 mm error in the MLC leaf pairs, can cause a dose error of more than 1.7 % (13,14).
Therefore, for small fields, intensity-modulated fields, and stereotactic treatments, the QA physicist must have available an efficient, robust series of procedures that can verify the accuracy of the MLC leaf dosimetry prior to the start of the course of treatment.
Previous QA Tests for MLC Accuracy
Discrepancies in FWHM Measurements
Unfortunately, due to the physics of small field dosimetry, the MLC position measured by the conventional FWHM is incorrect when applied to small fields. The results obtained during this study indicated that there was a need for a more suitable algorithm to reliably measure MLC leaf position for small fields.
Full details on the new algorithm developed are published in:
EPSM 2019, Engineering and Physical Sciences in Medicine. Phys Eng Sci Med 43, 297–462 (2020). https://doi.org/10.1007/s13246-019-00826-6
Reasons for failure of the conventional FWHM algorithm
Reasons for failure of the conventional FWHM algorithm
It’s important to understand why the conventional FWHM measurements fail in small field conditions. This is well described by IPEM Report Number 103, Small Field MV photon Dosimetry.
Overlapping aperture by penumbra
The conventional water tank ‘penumbra’ is described as the region of the beam profile spanning 80 % to 20 % of the beam profile edge, measured at 10cm deep in water. The penumbral width is governed by the size of the X-ray beam focal spot, the position and length of the jaw collimators, the position, and thickness of the collimators and MLC leaves and the lateral scatter of the X-ray beam for given field size.
If the radiation field defined by the collimators is less than or equal to the width of the penumbra, then the FWHM will be affected.
Lateral electronic disequilibrium
Electronic equilibrium is defined as:
the number of charged particles of a given type and energy leaving the volume equals the number of charged particles of the same type and energy entering that volume.
Failure to satisfy this condition is referred to as lateral electronic disequilibrium.
The EPID semiconductors are more sensitive to low energy scatter radiation. Small X-ray fields have less low energy scatter from the Flattening Filter as compared to the large fields which affect the resultant FWHM [24]. This means that neither the EPID dose calibration or the beam profile is exactly the same as the standard profile of Figure 2.
Small X-ray fields have less low energy scatter [4] as compared to the large fields which affect the resultant FWHM.
Radiation source occlusion
The X-ray beam profile changes as the collimators begin to obscure and block the primary radiation source, resulting in a reduction in the X-ray dose. Additionally, the profile distribution of photon energy exiting the collimator and MLC housing is not uniform. A mixture of primary, scattered and bremsstrahlung radiation produces an X-ray beam with higher energy photons in the central region than laterally and lower energy photons laterally. The energy spectra will change characteristics as the outer edge from the radiation source are blocked by collimation.
Algorithms investigated
The original design brief was to investigate algorithms used in computer vision imaging. The aim was to determine whether any of these algorithms were suitable for this application. Four different computer algorithms were tested to see whether the leaf positions could be accurately determined in small stereotactic type fields at sub-millimetre and sub-pixel levels.
Appendix A below provides a summary of the algorithm test methods.
QA Program for leaf position accuracy
The positional accuracy of all leaves (and backup jaws, if applicable) should be checked at least monthly. It is the responsibility of the qualified medical physicist to understand the MLC positioning system and decide which test is appropriate. The test should be performed at different gantry angles to detect any gravity‐induced positional errors.
An acceptable test includes a Picket Fence‐type test. Other tests that are tailored to the design of Elekta and Siemens MLC systems also exist (Hancock for Elekta and the Diamond jig system for Siemens). Leaves should move to prescribed positions to within 1 mm for clinically relevant positions (25).
Conclusion
Automatic and autonomous EPID-based methods were developed and applied to investigate sub-pixel algorithms for small MLC defined fields. The paper is essentially a negative result, illustrating the failure of simple derivative-based approaches for edge detection and for traditional FWHM methods of MLC leaf detection applied to small fields. The derivative-based approaches were found to be inadequate for very small fields due to overlapping penumbra shifting the point of inflection. A modified iso-intensity algorithm is presented as an improved method for MLC positional verification of small irregular shaped static fields, but further work would be required before it could be routinely introduced to a clinical setting.
References
Appendix A
Appendix A – algorithms investigated
Appendix A - algorithms investigated
The original design brief was to investigate algorithms used in the field of computer vision imaging. The aim was to determine whether any of these algorithms were suitable for this application. Four different computer algorithms were tested to see whether the leaf positions could be accurately determined in small stereotactic type fields at sub-millimetre and sub-pixel levels.
Detection of a collimator leaf position at better than the pixel level can be achieved using interpolation. This was found to be of more importance for older model EPID panels that have larger pixels than the newer models. For example, the aS500 panels have a (detector plane) pixel pitch of 0.784 mm (or approximately 0.5 mm at the iso-plane), which necessitates sub-pixel techniques when considering tolerances of the order of 0.5 mm.
Algorithms designed for image edge detection were chosen. They were:
Derivative Interpolation
This method was originally developed prior to the use of MLCs. It was designed to detect the edge of an X-ray field produced by the main collimators. Sudden changes in image intensity represent an edge within an image. Finding the extrema of the derivative of the image will localise these sudden changes.
The algorithm is a simple derivative interpolation (DI) consisting of two steps:
(a) Using the central difference of the EPID image to approximate the derivative of the averaged 1-D profile that spans a leaf pair, the extrema should correspond to the coarse location of the edges (at the pixel-level).
(b) Locate the sub-pixel position of the extrema via cubic spline interpolation of the five most adjacent data points. This is then assumed to be a suitable surrogate for the location of the radiation field edge.
Figure 7. The dashed region (green) shows the approximate location of the extrema in the derivative profile. Zoomed in on the left is a plot of the interpolation process in finding the maximum of the arrowed region.
The DI algorithm was also found to be imperfect when applied to small fields due mainly to penumbral overlap from the MLC leaf on the opposing bank.
Partial Area Effect
Figure 8. The output of pre-processing and segmentation algorithm applied to an EPID image of the dumbbell test shape formed using 24 field-defining leaf pairs of Varian2. The red cross marks the CAX determined from the reference images, black triangles mark the pixel location of the approximate centre of each leaf. Scale shows pixel values, which for our purposes have arbitrary units of intensity.
A slightly modified version of the Partial Area Effect algorithm, developed by Trujillo-Pino et al, 2013 [26], was applied to the MLC localise task of sub-pixel edge detection.
A short explanation of the method is as follows:
An edge pixel at a boundary contains information from two regions – (i) the region under the edge forming object and (ii) the region not covered by the edge.
The partial area effect hypothesis assumes that the signal intensity of the edge pixel is proportional to the areas of this pixel that is covered by the signal intensities from (i) the edge forming object and (ii) the area not covered by the edge – referred to as a partial area effect.
For each pixel containing an edge, the algorithm uses additional information from a surrounding pixel region to create a quadratic curve that predicts the edge boundary at sub-pixel level.
An excellent and more thorough description of the algorithm is described by its inventors in: https://doi.org/10.1016/j.imavis.2012.10.005 (Trujillo-Pino A, Krissian K, Alemán-Flores M, Santana-Cedrés D (2013), reference 24.
Figure 9. shows an example of Partial Area Effect (PAE) method applied to an EPID image of 10 x 10 cm2 MLC defined field. The middle image is a zoomed-in region highlighted by the dashed lines (green) and indicated by the green arrow which visualises the sub-pixel leaf edges and normals delineated by the algorithm. Further zoom of a section surrounding a single leaf edge is shown in the most right image.
Laplacian of Gaussian (aka Mexican hat filter)
The Laplacian of Gaussian (LoG) filter can detect a point where there’s rapid change in image intensity – otherwise known as an edge.
The method searches for zero-crossings in the second derivative. Derivative-based operators enhance noise, so the Laplacian is first convolved with a Gaussian for smoothing before being applied to the image.
Figure 10: Visualisation of the Mexican Hat filter.
The Szeliski method [27] of interpolating a linear function and locating the sub-pixel zero-crossing points were implemented for this algorithm to find the MLC positions in the EPID images.
Modified Iso-Intensity (empirical approach)
The modified iso-intensity (MI) algorithm applies an empirical approach to determine leaf position errors. The distance between the points of 50 % intensity is shown to be invalid for measuring small field sizes.
But, it may be possible to establish a different relationship between the intensity and the corresponding MLC position for each field size. The relationship between these parameters, using curve fitting techniques, were investigated on several Varian and Elekta linacs.
To detect leaf positional error, the MI algorithm takes the nominal planned positions as a starting input. For full details on the modified empirical approach developed during this research work, see the Physical & Engineering Sciences in Medicine Journal reference:
MLC Positioning Verification for Small Fields: A New Investigation Into Automatic EPID-based Verification Methods Joshua Hiatt, Godfrey Mukwada, Michael Barnes, Hans Lynggaard Riis, Du Huynh, Pejman Rowshanfarzad.
Joshua Hiatt, 12 September 2020
