Retrospective SPECT/CT dosimetry following transarterial radioembolization

Abstract Transarterial radioembolization (TARE) effectively treats unresectable primary and metastatic liver tumors through intra‐arterial injection of Yttrium‐90 (90Y) beta particle emitting microspheres which implant around the tumor. Current dosimetry models are highly simplistic and there is a large need for an image‐based dosimetry post‐TARE, which would improve treatment safety and efficacy. Current post‐TARE imaging is 90Y bremsstrahlung SPECT/CT and we study the use of these images for dosimetry. Retrospective image review of ten patients having a Philips HealthcareTM SPECT/CT following TARE SIR‐Spheres® implantation. Emission series with attenuation correction were resampled to 3 mm resolution and used to create image‐based dose distributions. Dose distributions and analysis were performed in MIM Software SurePlanTM utilizing SurePlanTM Local Deposition Method (LDM) and a dose convolution method (WFBH). We sought to implement a patient‐specific background subtraction prior to dose calculation to make these noisy bremsstrahlung SPECT images suitable for post‐TARE dosimetry. On average the percentage of mean background counts to maximum count in the image across all patients was 9.4 ± 4.9% (maximum = 7.6%, minimum = 2.3%). Absolute dose increased and profile line width decreased as background subtraction value increased. The average value of the LDM and WFBH dose methods was statistically the same. As background subtraction value increased, the DVH curves become unrealistic and distorted. Background subtraction on bremsstrahlung SPECT image has a large effect on post‐TARE dosimetry. The background contour defined provides a systematic estimate to the activity background that accounts for the scanner and patient conditions at the time of the image study and is easily implemented using commercially available software. Using the mean count in the background contour as a subtraction across the entire image gave the most realistic dose distributions. This methodology is independent of microsphere and software manufacturer allowing for use with any available products or tools.

LDM and WFBH dose methods were statistically the same. As background subtraction value increased, we found the DVH curves to become unrealistic and distorted.
Conclusion: Background subtraction on bremsstrahlung SPECT image had a large effect on post-TARE dosimetry. The background contour we de ned provides a systematic estimate to the activity background that accounts for the scanner and patient conditions at the time of the image study and is easily implemented using commercially available software. We found using the mean count in the background contour as a constant subtraction across the entire image gave the most realistic dose distributions.
Comparison of dosimetry from background subtracted SPECT images to image based dosimetry obtained via 90 Y PET images will be the subject of our next analysis.

Background
Transarterial Radioembolization (TARE) uses beta particle emitting microspheres to treat unrespectable primary and metastatic liver tumors. These Yttrium-90 ( 90 Y) containing spheres are injected locally around the tumor(s) where they will become lodged in the tumor microvasculature delivering 95% of the radiation dose within 11 days due to its half-life of 64.2 hours [1]. Mean range for radiation delivery is 2.5 mm with a maximum penetration depth of 11 mm, allowing for local irradiation of tumor cells while sparing healthy liver tissue [2,3]. During pretreatment planning, Technetium-99m macroaggregated albumin particles ( 99m Tc-MAA) are used to mimic the dispersion of microspheres upon treatment and are Page 3/18 considered an acceptable surrogate for the basis of dosimetry calculations [2,[4][5][6]. Although they act as an acceptable prediction for microsphere distribution there is still a need for post-TARE emission imaging for validation of microsphere placement, especially for patient safety [1,[7][8][9][10].
Upon completion of TARE treatment, a 90 Y bremsstrahlung single-photon emission computed tomography (SPECT) imaging study is acquired to identify any extrahepatic microsphere deposition [2,3,11]. Historically, these images are only used only for qualitative information due to their noisy nature but some groups have shown their potential in quantitative imaging upon extra image processing [2,12,13].
Many challenges are present using these 90 Y bremsstrahlung SPECT images quantitatively. One is due to the continuous bremsstrahlung spectrum with a wide photon energy range of 0-2.3MeV [1,14]. Another is a lack of distinct photopeaks that are routinely used in SPECT imaging [1,12,14]. The high-energy photons can penetrate through the collimator septa [1,14]. Just these few concerns lead to very noisy, low resolution, images.
There is a general consensus on the bene ts of accurate image based post-TARE dosimetry among researchers in the eld but a lack of information on best practices using commercially available software makes implementation across intuitions di cult [1,2,4,15]. Recently there have been advancements in understanding the promising nature of post-TARE 90 Y positron emission tomography (PET) image based dosimetry, however it needs to be noted that not all institutions may have the feasibility to obtain 90 Y PET images post-TARE. At such locations, it is still imperative to perform post-TARE dosimetry thus there is still a need to better understand how to best use bremsstrahlung SPECT images for accurate dosimetry.
In this study, we sought to create a way to address background noise on these post-TARE SPECT images on an easily implemented patient speci c basis and determine background subtraction effects on resulting dose distributions.
This study, to our knowledge, is the rst to look at a systematic estimate to the activity background that accounts for the scanner and patient conditions at the time of the image study and its effects on post-TARE SPECT based dosimetry.

Methods
Through an IRB approved retrospective study we created post-TARE image based dose distributions for 10 patients treated with SirTex SirSpheres™ 90 Y microspheres for hepatic malignancies at Wake Forest Baptist Medical Center. Each patient had a bremsstrahlung SPECT/CT (parameters shown in Table 1) utilizing a Philips Healthcare™ scanner following SIR-Spheres® implantation. The SPECT images were attenuation corrected with the CT data, which was the only scanner correction.
Using MIM Software SurePlan™ (MIM Software Inc., Cleveland, OH) the SPECT images were resampled to 3 mm resolution and contours were drawn for organs of interest based on the registered CT image. Dose distributions, as shown in Fig. 1, were created based on these bremsstrahlung SPECT images using SurePlan™ Local Deposition Method (LDM) and our own dose convolution method (WFBH), both applied at the voxel level. Through the process of creating dose distributions, the emission image was scaled based on actual delivered activity at the time of treatment. Within the LDM model, the normalization and dose calculation is based on counts found within the liver contour. We expanded this contour by 20 mm on all sides prior to any normalization in order to account for potential movement during imaging as well as if activity is located close to the edge of the liver. For WFBH, the normalization is based on the total counts found in the body contour. Table 2 shows the patient characteristics and treatment details of those included for this analysis. The main goal of this study was to understand how the noise within these bremsstrahlung SPECT images affects the resultant image based dose distributions. In order to try get a handle on eliminating the background noise in the image while not eliminating true counts, we created a background contour de ned to be the whole body contour minus the liver and lungs expanded by 20 mm. This contour was originally created on the CT image and transferred to the emission series where its statistics were recorded. Figure 2 shows an example of this contoured area, shaded in blue. In addition to dose distributions created using LDM and WFBH dose calculation methods on the original SPECT images, we created dose distributions after a constant background subtraction on the emission image was performed. We wrote an extension in Java (Oracle corporation et al., Redwood City, CA, USA) to run in MIM (see supplementary material for code) in order to easily subtract background on a patient by patient case. The extension prompts you to enter a background value (in counts) which will be subtracted from the entire emission image. If a given voxel value becomes negative due to the subtraction, it will set to zero (0). We de ned the following shorthand for the various corrected images we created based on the different background contour statistics being subtracted: Bkgrd = mean counts in background contour subtracted from all voxels, Bkgrd + SD = Bkgrd plus one standard deviation subtracted, Bkgrd + 2SD = Bkgrd plus two standard deviations subtracted.  As expected, Table 4 shows that as the subtracted background value increases, the max dose value increases for both dose calculation methods. This trend further demonstrates that a statistically de ned background subtraction is needed. No background correction can predict too low of a dose delivered while too high of a background subtraction will predict too high of a dose delivered. Ratio of max dose value obtained after background subtraction compared to its respective original image max dose value averaged over all patients The dose line pro les in Figure 4a demonstrate that absolute dose is effected by background subtraction, the maximum dose increases as the background subtraction value increases. When the dose line pro les are normalized to the maximum point in each respective method, we see that dose distributions are not spatially being effected for a given background level (Figure 4b). Also note in Figure 4b the low dose region shrinking as the background subtraction value increases. This demonstrates the potential of removing low dose delivered regions that could important to consider in future patient care. Through these dose pro les we see an agreement between LDM and WFBH dose calculation methods and Table 5 shows the agreement of the resulting maximum dose values. LDM yielded higher maximum dose values on average without background consideration but upon background subtraction the two methods give comparable ratios.  Figure 5 shows representative DVH curves for the listed liver dose distributions. We found that DVH curves for WFBH and LDM methods were not statistically different in 39 out of 44 curves through two tailed p-values (statistical signi cance deemed to be less than 0.05). In these 5 curves that were deemed statistically different, 3 were from the same patient and were for the resulting DVH curves after each background subtraction value. Figure 6 shows the DVH curves for this patient and as we see the LDM method is shifted below the WFBH method in these three cases. The other 2 statistically different curves were for just one set of curves in each patient. Across all three patients their treatment area is right at the edge of the liver contour.
In Figure 5-7 WFBH Bkgrd and LDM Bkgrd show an expected shape change for DVH curves due to the decrease in low dose and increase in high dose as well as healthy tissue sparing. Figure 7 shows DVH curves for the treated region, de ned to be the region that is up to 20% of the max activity. These curves demonstrate better dose coverage with Bkgrd subtraction value compared to no correction. At the same time, Bkgrd + SD and Bkgrd + 2SD DVH curves show that these background values are unphysically skewing the delivered dose distribution to less coverage at low dose values than no background correction hence an unphysical altering of dose distribution at these levels. With Bkgrd subtraction level we do not see this effect and it is deemed the largest value we can subtract without creating unphysical changes in the DVH curves.

Discussion
Although 90 Y PET imaging is thought to be superior for post-TARE dosimetric needs due to improved spatial resolution [16][17][18], it cannot be assumed that all institutions performing these treatments have access to such imaging modalities. Therefore, a better understanding of the dosimetric outcomes based on bremsstrahlung SPECT images is still necessary. By utilizing commercially available software in this study we are also enhancing the ability for post-TARE image based dosimetry calculations to become standard practice, something that is highly lacking in this eld currently due to the imaging complexities associated with 90 Y.
We explored one methodology of trying to eliminate false counts in the emission scan prior to dosimetry calculations. In theory, within the background contour we created, there should not be activity present so we used this area to help us identify on average what portion of counts within our organs of interest are true and which are false. Our rationale of expanding the liver and lung contour 2cm was to try to account for the potential breathing motion of the patient during imaging.
As the background value increased in magnitude, we found the overall max dose of the resulting dose distributions to increased, which was expected because absorbed dose within a voxel is calculated via the following equation: where A =activity, LSF = lung shunt fraction, T 1/2 = 90 Y half-life, E avg = average β-particle energy per disintegration (0.935 MeV), C voxel = counts within voxel, ΔV = voxel volume, ρ = tissue density, and C total = total counts within the patient [15]. The calculated dose in a given voxel is the voxel value divided by the total counts multiplied by a constant. So as the background subtraction level increases, the count total decreases, which is in the denominator, thus causing the calculated dose value to increase.
An assumption of the LDM model is that all the energy from β-particle decay is locally within the same voxel [15]. WFBH method uses a 5x5x5 scattering kernel to consider the effects on neighboring voxels. However, through our analysis in this study it appears this does not make a considerable difference in resulting dose distributions when using 90 Y Bremsstrahlung SPECT images. This is likely due to the mean range of β-particles in tissue being 2.5mm and using a resampled voxel size of 3mm in our emission scan. Availability of data and material: The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
Competing interests: BCT declares no con ict of or competing interests. WAD serves as an advisory board consultant for Sirtex Medical.   Visual representation of background subtraction on image Representative images showing for the same patient, the effects of our background subtraction technique. Each row represents a different subtraction value starting from the top row having no subtraction, second row mean value subtraction, third being mean + SD and last row is mean + 2SD. y-axis and an arbitrary location on the x-axis. Dark blue line represents WFBH method, green is LDM method, peach is WFBH Bkgrd, red is LDM Bkgrd, magenta is WFBH Bkgrd + SD, light blue is LDM Bkgrd + SD, yellow is WFBH Bkgrd + 2SD and grey is LDM Bkgrd + 2SD.

Figure 5
Liver DVH curves Showing the effect background subtraction has on the DVH curves for both dose calculation methods. We see dose method has little effect while background subtraction value has large effect. Figure 6 exampled of statistically different DVH curves We found that in a majority of cases these DVH curves between LDM and WFBH dose methods with and without background subtraction were not statistically different however this is an example of when this was not the case.