Image processing for IMRT
QA dosimetry
Mehran R. Zaini, Gary J. Forest, and David D. Loshek
Radiation Oncology Department, Marshfield Clinic, 1000 North Oak Avenue, Marshfield,Wisconsin 54449 U.S.A.
zaini.mehran@marshfieldclinic.org, forest.gary@marshfieldclinic.orgloshek.david@marshfieldclinic.orgReceived 15 October 2004; accepted 2 June 2005
We have automated the determination of the placement location of the dosimetry ion chamber within intensity-modulated radiotherapy (IMRT) fields, as part of streamlining the entire IMRT quality assurance process. This paper describes the mathematical image-processing techniques to arrive at the appropriate measurement locations within the planar dose maps of the IMRT fields. A specific spot within the found region is identified based on its flatness, radiation magnitude, location, area, and the avoidance of the interleaf spaces. The techniques used include applying a Laplacian, dilation, erosion, region identification, and measurement point selection based on three parameters: the size of the erosion operator, the gradient, and the importance of the area of a region versus its magnitude. These three parameters are adjustable by the user. However, the first one requires tweaking in extremely rare occasions, the gradient requires rare adjustments, and the last parameter needs occasional fine-tuning. This algorithm has been tested in over 50 cases. In about 5% of cases, the algorithm does not find a measurement point due to the extremely steep and narrow regions within the fluence maps. In such cases, manual selection of a point is allowed by our code, which is also difficult to ascertain, since the fluence map does not yield itself to an appropriate measurement point selection.
PACS numbers: 87.53.Xd, 87.66.Cd, 87.57.Nk
Key words: radiation therapy, IMRT, QA, image processing
I. INTRODUCTION
For every patient who undergoes intensity-modulated
radiotherapy (IMRT) treatment, our physics team performs a series of
quality assurance (QA) tests. These tests are to ensure proper
delivery of radiation dose to the appropriate geometrical sites
within the patient.
Using our
treatment-planning system, Philips Pinnacle®, we produce multiple
beams, each having varying radiation intensity patterns. Part of the
QA process involves placing the generated fluence patterns on a
flat-water phantom in the planning software. Dose maps at two planes
of 6 cm and 7 cm inside the phantom perpendicular to the beams are
computed. One of the two depths (7 cm) is for absolute dosimetry
verification at a single point, and the other one (6 cm) is for
comparing the computed isodose patterns with those of the irradiated
films. A spherical ion chamber (0.125 cm3) is used for the absolute measurements,
and Kodak EDR-2 films are used for relative dose maps. These
measurements are performed during the same irradiation session for
each beam. Absolute dosimetry with film is not as accurate and
reproducible as it is with ion chamber. Our tolerance for absolute
dosimetry accuracy is 5%. We are in the process of defining the
limits of acceptance for comparing the isodose patterns, since it
has so far been a qualitative comparison. For the isodose profile
comparison, we are developing image subtraction software, as well as
a system based on principal component analysis.
In addition to perpendicular dose
planes at two different depths for each radiation beam, a computed
transverse radiation map for the integral of all the beams is also
compared with a measured pattern using a film. This transverse image
is taken in either a cylindrical or thorax water phantom, which was
designed in our department.
The
entire IMRT QA procedure is time-consuming for the physics team. We
are devising a system that not only helps the physicists in
performing the IMRT QA tasks more efficiently, but also allows the
dosimetrist to be able to perform the IMRT QA measurement tasks with
the medical physicist's supervision. This system incorporates the
automation of some of the tasks that require trained eyes. One such
automation task involves the determination of the optimum placement
of a dosimeter (an ion chamber) within the intensity-modulated
field. A number of mathematical image processing techniques have
been designed and tested in order to achieve this automation. This
paper discusses the methods and results of the automated dosimetry
localization system.
II. METHODS
A. Planar dose maps
The topology of planar dose maps (PDMs) often has regions
with large gradients. Placing an ion chamber at such high-gradient
spots does not yield an accurate measurement. This is due to the
finite size of the detector, as well as the imprecision in physical
placement of the dosimeter at the intended position. If the detector
is misplaced in the high-gradient regions by 2 mm to 3 mm, the ion
chamber measurements will be different than the computed values by
more than our tolerance level of 5%. Furthermore, there are regions
within the PDM that have a low magnitude (<15% of the maximum
value), where the detector signal-to-noise ratio would be low,
leading to measurement errors. Absolute measurements at a low
magnitude area do not constitute a good test of the delivery system,
since the high-fluence regions deliver most of the dose to the
patient, and their dose deposition accuracy should be verified.
It is difficult and time-consuming
to find such a position visually by inspecting the radiograph-like
images of the PDMs. On the other hand, automatically determining the
location where an ion chamber measurement needs to be made is
nontrivial. Mathematical image-processing techniques allow the
determination of such a position. In short, a specific region is
identified based on its flatness, radiation magnitude, area, and the
avoidance of the interleaf spaces of the multileaf collimators.
B. Mathematical image processing
Figure 1 depicts the various steps involved in determining the most suitable location for making dosimetric measurements in patients' IMRT fields. MATLAB® by The MathWorks, Inc.,© was used for the algorithm development in this work. The steps in Fig. 1 are explained below:
Fig. 1. Mathematical image-processing flowchart |
||
Step 1: A Laplacian is used on the planar dose images(1):
(1) |
where f(x, y) is the
two-dimensional dose image as a function of spatial coordinates
x and y.
For the
sake of this work, we are using the word gradient instead of
curvature, despite the fact that a Laplacian creates a curvature
map. This oversight is justified by the method we used to extract
the desired information from the processed images. The kernel
associated with this Laplacian we have used to be convolved with the
image is
(2) |
Since the Laplacian is a second derivative operator, it will produce an abrupt zero crossing at an edge. The Laplacian is a linear, shift-invariant operator, and its transfer function is zero at the origin of frequency space. Hence, a Laplacian-filtered image will have zero mean gray level.(1) In our case, we are interested in a narrow band around zero, as explained later, and not the zero crossings themselves. An example of a PDM and its Laplacian-filtered image is shown in Fig. 2.
Fig. 2. The original PDM and its Laplacian-filtered image |
||
For a noise-free
image with sharp edges, a binary image formed by thresholding at the
zero-crossings of a Laplacian-filtered image will produce
closed-contoured regions of interest (ROIs). In our case, since we
seek gradient maps and no sharp edges normally exist, thresholding
at a clinically relevant value is appropriate. Furthermore, for an
image with noise, presmoothing with a Gaussian filter is the usual
practice. However, applying a low-pass filter prior to using the
Laplacian is not desirable for our application, since the
fluctuations in PDMs are desired variations and not unwanted noise.
Hence, we simply apply a single Laplacian to the images.
Step 2: The threshold value
that worked well was determined to be at ±0.02. This value is the
relative gradient value in the Laplacian applied image. Even though
this threshold value is adjustable, it rarely needs to be changed.
The tuning range for this threshold value is ±0.01 to ±0.025, where
± is applied within the code. Using this thresholding limit, a
binary mask is created. Upon thresholding, the image background gets
selected as an acceptable area as well. However, picking the
un-irradiated background areas as a ROI is problematic, since it
creates challenging issues at the subsequent processing steps. These
issues include (1) the elimination of the regions close to the edge
of the field that can occur in the subsequent image processing steps
and (2) the selection of the un-irradiated areas as a viable option.
Step 3: Removing the
background as a ROI cannot be made by simply flipping the binary
mask, since the binary mask is based on the gradient map, and the
edges have a low gradient. To do this, the perimeter of these
relatively flat regions is found. Then, upon labeling them (a MATLAB
function), we can suppress the background to become an unselected
ROI.
Step 4: Afterward,
upon applying a dilation and an erosion operation, we first reduce
the number of single streaks and small, unconnected regions and then
eliminate their remaining ones. General dilation is defined as(1)
(3) |
where D is the resultant binary image, B is the original image, and S is the set of points (x, y) such that if S is translated so that its origin is located at (x, y), then its intersection with B is not empty. In this case, S is a 3 × 3 unity matrix, which translates to a 0.56 cm2 area, since the pixel is 0.25 × 0.25 cm2. The algorithm is flexible with the pixel size, and its performance has been found not to be dependent on the dose grid sizes that have been tested (0.1 × 0.1 cm2 to 0.5 × 0.5 cm2).
Step 5: Similarly, general erosion is defined as(1)
(4) |
where E, which is the resultant binary image, is formed by eroding the original image B by S such that if S is translated so that its origin is located at (x, y), then it is completely contained within B. For the erosion operation, S is a 4 × 4 unity matrix, which translates to a 1.0 cm2 area. The sizes of S for the dilation and erosion operations were chosen based on the size of the ion chamber used for the measurements. The sizes of S for the two operations are adjustable by the user, but this is almost never required.
Step 6: These
dilation and erosion operations lead to the desired, relatively flat
ROIs being delineated and all other unwanted regions being
eliminated. The final ROIs get automatically identified by finding
their perimeters, and they are labeled with natural numbers.
Step 7: Since the
background gets labeled again, it is suppressed for the second time.
An example of a threshold-applied image and its outcome after the
dilation and erosion are shown in Fig. 3.
Fig. 3. Laplacian image thresholded (left) and after applying dilation and erosion operations |
||
Step 8: In
order to select the ROI that provides for both a reasonable
magnitude and area, an optimization scheme is designed. This
optimization process starts by modulating the ROIs by the original
PDM.
Step 9: Mean height
and area of the ROIs are computed to be used for the ROI selection.
The ROI area is weighted less than its mean height in the
optimization phase through a power parameter. This is because of the
magnitude differences between the areas and mean heights, and
because the heights of the ROIs are more important than their area
for the ROI selection process.
Step 10: The power
parameter was determined to be 0.2. The product of the rescaled mean
height and the area to the power 0.2 is the objective function to be
maximized:
(5) |
This value is adjustable by the user as well, but
in rare cases it needs to be changed to 0.25.
Step 11: The center of mass
(CM) of the selected ROI is the desired position for ion chamber
measurements. The CM computations include the details inside the
ROI. In general, the geometric moment of order p + q
for an image function f (x, y)is defined as(2)
(6) |
The center of mass is computed as
(7) |
where m00 =
N, the total number of pixels inside the region.
However, it is possible that this
CM falls into a narrow region of a ROI connecting two or more local
clusters. In such cases, it is necessary to breakup the ROI into
smaller regions. To do this, the smallest rectangle that encompasses
the ROI is found. The ratio of the points inside the ROI and that of
the bounding rectangle is computed. If this ratio is larger than
70%, it is reasonable to assume one main cluster, in which case the
CM is the proper choice. However, if this ratio is less than 70%, a
breakup of the ROI is warranted.
Step 12: Further
eroding and labeling the subregions accomplish this breakup. The
produced subregions need to be reevaluated against the previously
found ROIs to determine the optimum ROI to select.
Step 13: Finally, the CM of
the reselected ROI is chosen as the desired point of measurement.
III. RESULTS AND DISCUSSION
An example of a PDM image that depicts the automatically selected measurement point is displayed in Fig. 4. The selected point is accentuated in Fig. 4 for demonstration purposes only. It is apparent that determining the location of an appropriate measurement point visually is cumbersome. The coordinates of the identified point are passed to the user for the placement of the ion chamber in the phantom. The avoidance of the multileaf collimator interleaf spots is achieved using proper indexing of the results with respect to the actual leaf dimensions.
Fig. 4. A PDM with the selected point of measurement |
||
We have allowed for the possibility of failure of this code in extreme cases, in which cases the user can manually select a position for the ion chamber placement. In testing this algorithm in more than 50 cases, in only 5% of the beams one needs to resort to our manual point selection. The results of dose measurement agreements with Pinnacle predictions for a random sample of our IMRT patients are presented in Table 1. Even though it is possible to iteratively adjust the three parameters discussed earlier to achieve the point selection through this algorithm, the manual method is available if one does not wish to adjust any parameters. An automated parameter optimization algorithm is the focus of our future work.
| Table 1. Dosimetry results based on automated and manual point selections |
It seems reasonable
to assume that pixels larger than 0.25 × 0.25 cm2 would reduce the failure rate of this
algorithm, but then high-frequency intensity variations would be
missed. Smaller dose computation grids, on the other hand, not only
increase the PDM computation time; they would potentially lead to a
higher failure rate of this code to find a suitable dosimetry
localization point. Furthermore, a 0.125 cm3 ion chamber is a reasonable size of
measurement device. This is because larger ion chambers will reduce
the efficacy of IMRT QA, and the smaller ones will lead to enhanced
measurement noise and more susceptibility to positional accuracy.
The manual point selection
involves visually inspecting the promising areas, displaying the
magnitude of the point under the mouse and the mean gradient for its
eight-neighbor pixels, and picking a point based on its magnitude
and mean gradient. This selection is made by deciding on a point
based on the compromise between a high magnitude and a low immediate
neighbor gradient. Such a determination is made by observing a few
promising regions. It is not based on an exhaustive search, since a
complete search takes more than a few minutes per beam. The
automatic code of this paper was developed precisely because a
manual and visual point selection is time-consuming, it is not as
accurate, and it might not lead to the best measurement point.
We attempted to determine a
measurement point using the tools available in Pinnacle. This
process alone took more than 30 min, whereas our automated code
achieves the same outcome in 2 min for all the IMRT beams of a
patient. The process of using the Pinnacle tools involved taking the
IMRT plan to a water phantom, repositioning the computation points
of interest, reorienting the beams, recomputing the dose for all the
beams, and creating the isodose lines in the plane of interest.
Using the Pinnacle tools takes considerable training of the operator
to get consistent results. One of the goals of this algorithm is to
make the process less subjective. On the other hand, the planar dose
computation window of Pinnacle does not allow one to create isodose
lines on it. Furthermore, the window/level tool does not help one to
inspect the PDM, in the manner it is displayed on the planar dose
computation window. Henceforth, using Pinnacle for the purposes of
this work is not appropriate.
IV. CONCLUSION
The automated dosimetry localization technique
discussed here simplifies and expedites the IMRT QA process for the
medical physicist. It also allows the dosimetrist to perform the
IMRT QA procedure almost independently. The clinical implementation
of this system simply involves reading the PDM produced by Pinnacle
into the code, and letting the automated code show you the
measurement point. If the selected point is in a low magnitude
region (as seen visually by its gray-level magnitude compared to the
rest of the image), one can adjust one or two of the parameters to
arrive at a more reasonable point. In 95% of the cases, such
parameter adjustment is not necessary. If one does not want to
adjust the parameters, one has the option of selecting the point
completely manually by just observing the magnitude and the local
gradient under the motion of the mouse. Both methods establish
suitable dosimetry verification points where measurements agree with
computer predictions within our tolerance limit of 5% for correct
IMRT plans.
We are working on
automating more pieces of the IMRT QA process, such as quantitative
film analysis, the creation of the lookup table for film dosimetry,
and the incorporation of these pieces into our record-and-verify
system. In addition, we are in the process of creating a more robust
optimization module, where the parameter space is also searched
automatically.
ACKNOWLEDGMENTS
The authors thank Marshfield Clinic Research Foundation for its support through the services of Alice Stargardt in the preparation of this manuscript.
REFERENCES
- Castleman KR. Digital image processing. Upper Saddle River (NJ): Prentice-Hall; 1996.
- Theodoridis S, Konstantinos K. Pattern recognition. 2nd ed. San Diego (CA): Elsevier Academic Press; 2003.
© 2005 Am. Coll. Med. Phys.