Automated MV markerless tumor tracking for VMAT

Abstract

Tumor tracking during radiotherapy treatment can improve dose accuracy, conformity and sparing of healthy tissue. Many methods have been introduced to tackle this challenge utilizing multiple imaging modalities, including a template matching based approach using the megavoltage (MV) on-board portal imager demonstrated on 3D conformal treatments. However, the complexity of treatments is evolving with the introduction of VMAT and IMRT, and successful motion management is becoming more important due to a trend towards hypofractionation.

We have developed a markerless lung tumor tracking algorithm, utilizing the electronic portal imager (EPID) of the treatment machine. The algorithm has been specifically adapted to track during complex treatment deliveries with gantry and MLC motion. The core of the algorithm is an adaptive template matching method that relies on template stability metrics and local relative orientations to perform multiple feature tracking simultaneously. Only a single image is required to initialize the algorithm and features are automatically added, modified or removed in response to the input images. This algorithm was evaluated against images collected during VMAT arcs of a dynamic thorax phantom. Dynamic phantom images were collected during radiation delivery for multiple lung SBRT breathing traces and an example patient data set. The tracking error was 1.34 mm for the phantom data and 0.68 mm for the patient data.

A multi-region, markerless tracking algorithm has been developed, capable of tracking multiple features simultaneously without requiring any other a priori information. This novel approach delivers robust target localization during complex treatment delivery. The reported tracking error is similar to previous reports for 3D conformal treatments.

1. Introduction

Addressing tumor motion during radiotherapy treatment is integral to ensuring the target receives the prescribed dose. The amplitude of motion is dependent on anatomical location, and can be relatively large for lung cases (up to 2cm) (Ekberg et al.; 1998; Seppenwoolde et al.; 2002; Keall et al.; 2006), and also for prostate (Kitamura et al.; 2002), adrenal (Katoh et al.; 2008) and liver (Liang et al.; 2018). With the introduction of intensity-modulated radiation therapy (IMRT), highly conformal treatments with steep dose fall-offs can be delivered more easily. However, the limit to faithfully delivering the planned treatment dose while reducing dose to healthy tissue is still constrained by the gross tumor volume (GTV) expansions for tumor motion. The emergence of advanced modalities like hypofractionated stereotactic body radiation therapy (SBRT) and adaptive radiotherapy provides further impetus for robust motion management.

A variety of strategies have been developed to address tumor motion, generally falling into two categories: one is to limit motion to an acceptable range and adjust treatment margins accordingly, the other is to track and respond to the motion in real time. The former includes techniques such as abdominal compression, although the success of this technique for abdominal cancers is inconsistent (Van Gelder et al.; 2018; Lovelock et al.; 2014; Eccles et al.; 2011) and there are studies showing counterproductive results cases for lung SBRT (Mampuya et al.; 2014).

Active strategies rely on in-treatment tumor localization and corresponding adaptation. The position monitoring can be done through the use of an external or internal surrogate, or through some form of imaging (kV, MV, MR or ultrasound (Mostafaei et al.; 2018)). These methods of tumor tracking have been combined with beam gating (Berbeco et al.; 2005), or multi-leaf collimator (MLC) tracking (Keall et al.; 2000) to direct the correct dose to the target. However, the use of a surrogate either assumes or requires some modelling of the correlations between the surrogate and tumor motion, introducing additional uncertainties. Furthermore, the placement of external surrogates must be reproducible which is not guaranteed, and the implantation of internal surrogates, e.g. fiducials, can be limited by patient concerns (Scher et al.; 2019), indicating that a general motion management strategy relying solely on surrogates may not be attainable.

The use of imaging modalities to directly track tumor motion somewhat addresses uncertainty in modelling, however both ultrasound and kV imaging require some assumptions about tumor motion due to their relative orientations with respect to the beam. Utilising the MV on-board portal imager for tumor tracking removes these considerations entirely, while also adhering to AAPM goals to reduce imaging dose (Murphy et al.; 2007). The electronic portal imaging device (EPID) mounted on treatment linear accelerators collects images generated using the MV treatment beam. These beams-eye-view (BEV) images are inherently of lower quality compared to kV images due to the physical processes that occur at higher energies. The higher noise and reduced contrast of MV images makes for a more challenging tracking environment, however previous work by our group demonstrates that tracking with these images is feasible to within 2mm for 3D conformal lung treatments (Rottmann et al.; 2010) and continued development of EPID technology (Rottmann et al.; 2016) to improve photon detection efficiency indicate a promising future for MV tracking (Yip et al.; 2017).

In this paper we present a multi-region algorithm for markerless beams-eye view tracking to enable tumor tracking during volumetric modulated arc therapy (VMAT) deliveries. This algorithm maintains the same advantages as previously discussed (adaptability to changing or irregular breathing patterns, robustness to target deformations) while removing the requirement of a training period and introducing tracking while the MLCs are in motion and/or the gantry is rotating.

2. Method

2.1. Algorithm Overview

The tracking algorithm is built to automatically detect features of interest and track those features while visible within the aperture. To overcome the challenges of markerless tracking during treatments which include MLC and gantry motion, each feature that is identified for tracking is described by a cluster of templates generated along lines of high local variance. Template matching is performed on subsequent images using the calculated normalized cross-correlation (NCC) on the variance map of MV images to locate the template (image segment) on the image. The use of a cluster of templates serves two purposes; first it effectively encodes the shape of the feature, and second the reliance of only local information allows the tracker to adapt when features come in and out of view, or are lost, without impact on the tracking of other features. Tracking of a feature requires successful template matching of a majority of templates within the cluster, and preservation of the feature shape. An overview of the algorithm steps is shown in figure 1.

Figure 1:

Flow diagram of the markerless tracking algorithm.

2.2. Image Processing and Aperture Masking

To prepare each image for tracking, image smoothing and aperture masking are performed. Smoothing is performed through the application of a box filter and is required to reduce noise in the image and as well as limit the appearance of very small, high intensity features which may impede tracking. The aperture is masked through an adaptive thresholding algorithm, which separates foreground and background. The threshold is found through an iterative procedure which calculates the mean pixel value above and below the initial threshold, and adjusts this by the absolute difference in the calculated mean values until convergence. This algorithm was found to be successful for even complex apertures if the jaw positions are used to select a subsection of the image for processing.

The masking algorithm successfully detects the edge of the MLCs visible in the image, however further processing is required before tracking can be performed. The tracking algorithm first identifies regions of high variance and then tracks those features. The region close to the MLC edge, and hence close to the masked aperture edge, has a very high gradient in the image and so, without proper consideration, could result in inadvertent tracking of MLCs. To mitigate this, the aperture mask is narrowed to exclude the boundary gradient.

Template matching based tracking is limited to images where the aperture is open enough for templates to be generated (area > template size + exclusion region) so if this criteria is not met, the image is not processed and the algorithm progresses to the next. In addition, the number of tracked features is adapted to the aperture area.

2.3. Feature Detection

A template matching based method relies on uniquely identifiable features in the templates to perform successful tracking. The initial selection of these features is done by ranking regions of high local texture. The local texture is computed by means of a local variance filter (Haralick; 1979) where the pixel value is set to the calculated local variance, as described in equation 1, where $Q_{p_i}$ is a small region centered around the pixel p_i and containing M pixels. ${\overline{Q}}_{p_i}$is the mean value of all pixels within $Q_{p_i}$. Once the highest texture values have been located (while satisfying minimum separation requirements from other features and the aperture boundaries) they are ranked by a score based on the texture value and the uniqueness of the feature parameterized by the autocorrelation value.

$${Var\left(Q_{p_i}\right)=\frac{1}{M-1}\sum \left(Q_{p_i}(x)\right)-{\overline{Q}}_{p_i})}^2$$ (1)

In comparison with previous work (Rottmann et al.; 2010), a major change has been to absorb the pre-filtering of features into the main tracking procedure. The motivation for this adjustment is due to the dynamic nature of treatment types that are being targeted, which may not provide a stable and consistent set of images for pre-filtering. For the current implementation, if a feature is found to be static it is removed from the set of tracked features, and another is found to replace it using the same procedure outlined above.

2.4. Template Creation

Previous versions of this tracking algorithm relied on the preservation of relative orientations of the features to make a determination on whether the feature had been reliably tracked or not. This approach works well for 3D conformal treatment plans but faces challenges for treatments with MLC or gantry motion. To resolve this, we generate a set of templates (each 21×21 pixels in size) in the region around each feature identified for tracking. The number of templates generated as part of a single cluster is 7, chosen to minimize computational cost while improving the stability of feature tracking. The locations of these templates are determined by the path of the high local variance. This configuration allows us to infer the continued successful tracking of the feature based on the normalized cross-correlation values and relative orientation of these local templates, among other criteria, and also encodes information on the shape of the feature based on their orientation.

2.5. Template & Feature Tracking

Each feature is defined by a set of templates, the absolute position and relative orientations of which provide information that is used to determine whether or not a feature has been successfully tracked. The first step of this process is to calculate the new template positions using the 2D normalized cross-correlation of the templates with the current image. Next some selection criteria are applied to filter templates that have not been matched correctly, including a threshold on the normalized cross-correlation (0.5 for each template, 0.65 for the cluster average) and flagging templates that are static, where static templates are identified that have not moved more than one pixel per image for 5 images. An analysis is then performed of the relative orientations of the passing templates as follows. When the cluster of templates describing a single feature is created, a matrix encapsulating the position differences in each dimension between templates is calculated. On subsequent images, the absolute difference between the original matrix describing the relative orientations and the current matrix is examined.

Each element is inspected in sequence, and if this element is above a predefined threshold (0.35 pixel) the contributing template is removed from the matrix and the difference comparison is restarted until all values are less than the threshold, resulting in a subgroup of templates that preserve their relative orientations. Subsequent to this, a recombination check is performed to see if any templates removed in the first run-through of comparisons also form a subgroup that would pass the relative orientation check. The reason for this is that there is no guarantee that the resulting group from the first run through of the relative orientation check successfully track the feature, so we introduce the recombination stage to address this possibility. The score of the subgroups is calculated as a weighted combination of the normalized cross-correlation value of the templates, divergence from the original relative orientations and the average stability of the templates in that group (how often they have contributed to successful feature tracking). Introducing a score at this stage ensures that features that are tracked are done so with a high level of confidence, and promotes the use of stable templates over the sequence of images.

2.6. Template Updates and Lost Features

The final step of the tracking algorithm was introduced to improve tracking stability during VMAT treatments, where the perspective of the anatomy visible in the image is changing due to the rotating gantry. For features that are tracked successfully, with a high fraction of templates contributing to the matched position, the tracked templates are updated with the appropriate subsection of the current image. The cases where this update is performed are restricted to features where the confidence in the tracked position is high, avoiding the potential for templates to migrate to track an incorrect part of the anatomy.

2.7. Evaluation

To correctly evaluate the performance of the tracking algorithm, a set of images where the position of the tumor is known is required. Details of the equipment used to generate this set of images are provided in the following sections. The output of the tracking algorithm is backprojected to the isocenter plane and directly compared to the known tumor positions, projected onto the same plane. Histograms of the position differences in the x and y direction of the plane are generated and then fit with a normal distribution. The mean and width of the fit output are combined in quadrature to generate the tracking error in each direction of the 2D plane (referred to as x and y). These errors are combined, again in quadrature, to generate the final value for the tracking uncertainty.

Due to the nature of the treatment plan being delivered (VMAT), there may be periods of image acquisition where the tracking algorithm is not able to report a reliable tracked position. To quantify this a metric referred to as the tracking efficiency, the percentage of images reporting a successful tracked position, is evaluated and reported for each treatment field.

3. Experimental Setup

3.1. Phantom Data

To quantify the performance of the tracker a set of images is required with truth data for comparison with the tracker output. To create this dataset, images were collected of a dynamic thorax phantom (CIRS, Norfolk, VA, USA) containing a 2cm spherical tumor phantom (figure 2) in cine mode by the EPID imager mounted on a Varian TrueBeam LINAC (Varian Medical System, Palo Alto, USA). Images were collected directly from the LINAC using the iTools Image Capture software (Varian Medical System, Palo Alto, USA) and stored in ‘xim’ format. The acquired image size is 1280 × 1280 pixels at intervals of approximately 0.1 seconds. and the SID (source to imager distance) is set to 150 cm for the phantom images and 180 cm for all patient images, resulting in a spatial resolution of 0.219 mm and 0.183 mm respectively. Images were collected during treatment at intervals of approximately 0.1 seconds.

Figure 2:

Cross-sectional views from the CBCT of the thorax phantom.

The delivered treatment plan was adapted, without reoptimization, from the SBRT plan of a lung patient and consists of 3 VMAT arcs, covering 160 degrees each, delivering a total dose to the target volume of 1800 cGy. A table summarising the parameters of each arc is shown in table 1.

Table 1:

Parameters of the delivered treatment plan.

Field	Beam Energy	Gantry Rtn	Coll Rtn	Couch Rtn	MU
1	6x FFF	20 CW 179	90	0	2479
2	6x FFF	179 CCW 20	5	0	2552
3	6x FFF	20 CW 179	355	0	2427

To demonstrate the performance of the tracker in a more realistic scenario, the programmed motion of the phantom was modelled on patient breathing traces collected using an external sensor placed on the patients chest during treatment (Varian RGSC system). These data were collected in the clinic during lung SBRT treatments and the traces from 3 patients were used to drive the motion of the phantom in turn. The phantom was set up to replicate the positioning of a patient on the couch during treatment, so the displacement axis of the internal cylinder of the phantom generates the superior-inferior (SI) motion, and the rotation axis of the cylinder generates both the left-right (LR) and anterior-posterior (AP) motion. To map the patient breathing trace onto these two degrees of freedom, the amplitude of the target SI motion is scaled to the extent of motion observed on the patients 4DCT. The LR trace is similarly scaled by the amplitude of the tumor LR motion and used to program the rotation of the cylinder. The resulting motion extracted from the 3 breathing traces are shown in figure 3.

Figure 3:

Programmed displacement (top row) and rotation (bottom row) of the dynamic thorax phantom mapped from the breathing traces of three lung SBRT patients.

3.2. Patient Data

To illustrate the portability of the tracker from phantom to patient data, images were collected during a lung SBRT treatment on a Varian TrueBeam LINAC (Varian Medical System, Palo Alto, USA). The lesion is located in the upper right lung, and is approximately 3.5 cm, 3 cm and 3 cm in diameter in the SI, LR and AP dimensions (figure 4). The patient’s 4DCT showed a range of motion of approximately 8mm, 3mm and 6mm in the SI, LR and AP directions respectively. The images were collected in the same manner as for the phantom images; the MV panel was extended during treatment delivery and images were collected in cine mode by the EPID imager using the iTools Image Capture software (Varian Medical System, Palo Alto, USA).

Figure 4:

Beam setup as seen in the planning software for a single treatment field for lung SBRT.

4. Results

The markerless tracking algorithm was run retrospectively on MV images collected during the delivery of the previously described treatment plan to the dynamic thorax phantom. Three sets of images were recorded, corresponding to the delivery of the treatment plan to the phantom during the three separate patterns of programmed motion. To extract a quantitative metric for the tracking uncertainty, the tracked displacement of features from the images must be compared to the expected motion. This comparison is performed at the isocenter plane orthogonal to the central axis of the treatment beam. The 3D motion generated by the displacement and rotation of the cylinder within the dynamic phantom is projected to a pixel position on the MV imager, and then backprojected to the isocenter plane, allowing for a direct comparison of displacement values. An example of the tracker output in comparison to the projected phantom motion is shown in figure 5.

Figure 5:

Comparison of tracker output against the programmed phantom motion in the frame of reference of the MV imager projected to the isocenter plane. Each track refers to a tracked feature.

Due to the interference between the aperture size, position and shape, gantry angle and the motion of the tumor phantom itself the average tracking efficiency(number of images out of the series, roughly 1200 images, which generated a tracking output) is low and was found to be 16%, 8% and 8% for the first, second and third treatment fields respectively. There are parts of the image sequences in which the tumor is not visible, either due to mismatch of the aperture position and the tumor position at that time, or the aperture is not large enough for the tumor to be visible. Thus when tracking is re-established, the displacement centre for that track is set to zero. In order to make a comparison between external information and the tracked positions, the origins of the two coordinate systems must be aligned. To do this we seed the initial track offset by the value of the function describing the tumor motion at that time. The data points used to align the coordinate origins are not used in the evaluation.

The tracking error is evaluated using the method described in 2.7. This was initially performed separately for each delivered treatment field to check for any trends in the uncertainty based on the phantom motion and was performed for the full data set together to measure the total uncertainty (figure 6).

Figure 6:

Difference between the tracker output (displacement from original position) and displacement function in the x (a) and y (b) direction of the image.

The tracking errors resulting from the fit to the difference histograms are summarised in table 2. For each treatment field the total error is less than 1.5 mm, and the total tracking error was found to be 1.34 mm.

Table 2:

Tracking error summary table. The tracking error is broken down for each patient trace used, field and dimension on the imager plane at ISO. The 95% error is the value under which 95% of the data points fall and the % tracked describes the percentage of images that reported a successful track.

	Error [mm]
	Trace 1			Trace 2			Trace 3
Field	1	2	3	1	2	3	1	2	3	Total
X	0.52	0.98	0.82	0.69	1.20	1.16	0.77	1.00	1.14	-
Y	1.18	1.01	0.50	1.2	0.93	0.68	1.00	0.72	1.41	-
Total	1.29	1.41	0.97	1.38	0.97	1.34	1.26	1.23	1.81	1.34
95% Error	2.71	2.62	1.88	2.54	2.43	2.52	2.46	2.38	3.53	2.64
% Tracked	15.7	9.3	7.6	15.7	7.6	6.7	17.5	8.8	8.9	10.9

4.1. Patient Data Example

We applied the tracking algorithm to MV image data acquired during a lung SBRT treatment. A section of 60 images was chosen, sufficient to cover a full breathing cycle. Two example MV images are shown in figure 7. The example images (left/middle of figure 7) show two MV frames separated by approximately 1 second. Each tracked feature is indicated as the cluster of template positions, which are highlighted as magenta or red depending on whether the individual templates have been successfully tracked or not. The cyan arrows indicate the position relative to the original template location. Tracking was successful on 80% of the images, with majority of cases where tracking was lost being only for a single image, and the maximum period tracks were lost for was over 3 consecutive images (approximately 0.3 seconds). For this example, extracting the delivered dose from the image xim files and summing over images where a track was not reported, the number of MU delivered in that period was 15 MU.

Figure 7:

Screenshots of tracking output on patient data approximately a second apart (left and middle). The RHS plot is an example of tracking on the phantom data. The crosses indicate the original feature position (magenta if tracked, red if previously tracked but lost on the current image) and the cyan arrows indicate the current tracked position.

To estimate the tracking error, the output was compared to manually tracked positions. As the features are allowed to move independently, this was performed on a feature-by-feature basis. To track manually, a dedicated template is created at the feature location on the image where the feature is generated by the tracker. An expert then submits a match position for the template on subsequent images, until the template can no longer be matched successfully. The difference between the manually tracked position and the output of the tracker is computed. The histogram of these difference values is used to calculate the tracking performance. The same procedure as for the tracker output was followed to evaluate the tracking performance, and the fit to the histograms to generate the error in the x and y plane of the MV panel are shown in figure 8. The resulting tracking uncertainty from the evaluation of these 60 images was 0.68 mm.

Figure 8:

Difference between the tracker output (displacement from original position) and manually tracked positions in the x (a) and y (b) direction of the image.

5. Discussion

Tumor tracking based solely on MV images is a challenging task. Our approach to this challenge is conservative, with mechanisms in place to ensure tracking is only reported when there is confidence that the feature is tracked successfully. This reduces the tracking efficiency in terms of the number of images out of the total collected during the treatment field delivery that reported a tracked feature. In particular, for the images collected for this study, the imaging table installed on the couch creates a lattice effect on the MV images that moves as the gantry rotates. Creating a template on these grid lines introduces the possibility that the tracking would follow the gridlines rather than the intended target. To address this, we have introduced masking of vertical lines in the images, which reduces the tracking area available. This issue will be resolved in the future by a new imaging couch that does not produce a grid pattern on images.

Contributing to the low number of images that successfully produce tracks is the phantom itself, which is relatively homogeneous compared to real patient anatomy. This results in templates created mainly on the boundary between the phantom tumor and the surrounding material. Whereas tracking on real anatomy benefits from the texture of adjacent tissues, providing a larger region during MLC motion will present more availability of unobscured, trackable regions.

In comparison to the tracking error observed from the phantom images, our initial evaluation using a subset of images from an SBRT lung treatment demonstrates a lower value for the tracking error. We believe this is due to the type of phantom used for the study, in particular the shape of the tumor phantom. The size of each individual template is 19 × 19 pixels, which is 3.47 mm in the isocenter plane. The tumor phantom is a sphere of 2 cm diameter. Templates are created along the edge of the tumor phantom insert, because they generate smooth lines of high variance. The smoothness of these lines and the homogeneity of the material results in the generated templates being centered on features that are not fully unique, and the templates can match on regions slightly above or below the initiated point on the edge of the sphere with high confidence, potentially resulting in an increased tracking uncertainty.

Currently our tracking algorithm is applied retrospectively to series of images, and tracking on each image is completed in approximately 0.2 seconds, longer than required for running on real-time images (collected each 0.1 seconds, although this does not include any additional time required for transferring the image or accessing it through the API). However, there is room to improve by streamlining some of the processing, for example, when performing the normalized cross-correlation calculation for template matching some calculations are included to ensure any MLC boundaries and the image region beyond this to not contribute to a match. These calculations could be moved up a level in the structure so they are performed on a per-image basis. In addition, profiling the code shows there are instances where functions are repeated, which could be solved by passing information rather than recalculating. Finally, the computer being used contains an Intel ES-1630 CPU (2014), which is outperformed by CPUs found in newer PCs. Considering these factors we believe the algorithms will perform tracking at a rate fast enough to implement real-time tracking.

A review of the literature reveals a range of approaches to tracking on MV images including single template matching (Arimura et al.; 2009) and registration of simulated MV digitally reconstructed radiographs (DRRs) (Rozario et al.; 2018) which track tumor position with an uncertainty of 1.47±0.60 mm and 0.98 mm respectively, and a machine learning approach (Tang et al.; 2009) to detect when the tumor is outside of the aperture. The tracking uncertainty reported here is comparable to these values, and also to our groups previously published algorithm (Rottmann et al.; 2010) as well as markerless tracking algorithms utilising alternative sources of information, for example, on-board kV images (Hazelaar et al.; 2018).

Furthermore, to our knowledge this is the first presentation of markerless tracking using MV images collected during the delivery of a VMAT treatment. The demonstration of this tracking algorithm on phantom and an example patient data shows that this approach is feasible for complex delivery techniques with low tracking uncertainty. The previous functionality of the tracking algorithm has been preserved, and indeed should be improved by the template update process in reaction to inflating and deflating lungs. As an aside, by introducing feature recognition at the template generation stage, this algorithm can also be used for fiducial tracking. Further validation on a large clinical data set is required to fully explore the advantages and limitations of this method. Once the verification against the clinical data has been performed, an investigation towards clinical implementation of the tracking algorithm can be explored, building on previous work performed by our group investigating the use of a combined tracking and prediction algorithm to perform MLC tracking (Rottmann et al.; 2014).

6. Conclusion

An algorithm has been developed to perform markerless tumor tracking on MV cine EPID images collected during VMAT treatments. The tracking uncertainty of this algorithm was evaluated with a dynamic thorax phantom programmed with motion derived from patient traces. The reported tracking error is 1.34 mm, comparable to previously reported tracking errors by our group on 3D conformal treatments. Further studies will confirm the accuracy on a large clinical data set.