A phantom based evaluation of the dose prediction and effects in treatment plans, when calculating on a direct density CT reconstruction

Abstract In radiation therapy, a Computed Tomography (CT) image is needed for an accurate dose calculation. To allow such a calculation, the CT image values have to be converted into relative electron densities. Thus, standard procedure is to calibrate the CT numbers to relative electron density (RED) by using a phantom with known composition inserts. This calibration curve is energy and CT dependent, therefore most radiotherapy CT acquisitions are obtained with 120 kVp, as each tube voltage needs an additional calibration curve. The commercially available DirectDensityTM (DD) reconstruction algorithm presents a reconstruction implementation without any dependence on the tube voltage. In comparison, it allows a calibration curve that is directly proportional to the RED, reducing the need of more than one calibration curve. This could potentially optimize CT acquisitions and reducing the dose given to the patient. Three different phantoms were used to evaluate the DirectDensityTM algorithm in simple and anthropomorphic geometries, as well as setups with metal implants. Scans with the DD algorithm were performed for 80, 100, 120, and 140 kVp. As reference a scan with the standard 120 kVp scan was used. Radiotherapy photon plans were optimized and calculated on the reference image and then transferred to the DD images, where they were recalculated. The dose distributions obtained this way were compared to the reference dose. Differences were found mainly in pure air and high density materials such as bones. The difference of the mean dose was below 0.7%, in most cases below 0.4%. No indication was found that the algorithm is corrupted by metal inserts, enabling the application for all clinical cases. This algorithm offers more variability in CT parameters for radiation therapy and thus a more personalized image acquisition with a high image quality and a lower dose exposure at a robust clinical workflow.


| INTRODUCTION
In radiation therapy, the Computed Tomography (CT) images provide the essential patient geometry and electron density information needed not only to delineate the tumor and organs at risk but also to calculate dose. Beforehand, a relation between CT numbers and relative electron density (RED) or mass density (MD) has to be established in the form of a calibration curve in the treatment planning system (TPS) as described by e.g., Schneider et al. 1 . In the common case, the CT pixel values are given in Hounsfield units (HU) and depend on the tube voltage. Thus, a calibration curve for each tube voltage is needed. 1,2 Some clinics feature more than the standard 120 kVp calibration curve e.g., an additional 80 kVp curve for pediatric patients. This potentially reduces the CT dose for the patient or offers a higher contrast-to-noise ratio but increases the amount of work while the workflow robustness is decreased by the additional selection of the calibration curves. Other clinics only implement the standard 120 kVp curve for all scans ignoring the potential benefit of a higher signal-to-noise ratio or a decreased dose exposure that could be achieved with an optimized tube voltage. [3][4][5] The commercially available reconstruction algorithm 'DirectDensity' (DD) (Siemens Healthcare, Erlangen, Germany) promises a feasible workflow as it constructs the RED information directly from the raw CT data. 6 DD-CT images obtained this way are independent of the tube voltage once the calibration curve is implemented in the TPS.
Van der Heyden showed in a patient study for 33 patients that the mean dose difference between HU based dose calculated plans and those calculated on the DD image sets were found to be smaller than 1%. 7 In this study, the relative mean dose difference was evaluated for the planning target volume (PTV) and the organs at risk (OAR), as well as the relative difference in the volume receiving 95% of the prescribed dose were evaluated. These differences were also found to be less than 1%. These results are in agreement with the evaluation presented by Ritter. 6 Nevertheless, both works do not differentiate for cases that provide high uncertainties in imaging and dose calculation, such as air or metal cavities.
The present study analyzes the effect of the DD images on the dose distributions for a phantom with materials with high, medium and low density, exemplary for the density range present in the human body. As analyzing tools, a simple dose difference is accompanied by dose profiles, allowing a distinct analysis where the differences occur.
Additionally, CT scans with metal implants are compared as well as the combination of the DD algorithm with a metal reduction algorithm, 8 to evaluate if DD can be used in all clinical cases. Otherwise, for metal implants a standard conversion curve would be needed, revoking the easy setup and robust usability when implementing the DD algorithm.
Finally, the potential CT dose reduction benefits that could be utilized when dose acquisition parameters (e.g., the tube voltage) are optimized for image quality, are presented.

2.A | DirectDensity reconstruction
To understand the procedure behind the DirectDensity images and understand problematic setups and cases, a brief introduction of the algorithm is given. A more detailed description is given by Ritter. 6    The connection between the image values of the final DD image and the RED is given in eq. (1). 6

2.B | Density calibration
To obtain the RED information from the CT scan, calibration curves for 80, 100, 120, and 140 kVp tube voltage were generated with the Gammex 467 Tissue Characterization Phantom (Gammex, Giessen-Allendorf, Germany) for the HU and the DD reconstruction. All scans were taken with a constant tube current of 250 mAs on the SOMATOM Confidence (Siemens Healthcare, Erlangen, Germany).
The reconstruction with 3 mm slice thickness was executed with the B40s and E30s convolution filters. For the calibration, the mean CT value of each tissue mimicking insert was plotted against the RED of the material. The mean CT value for each insert was obtained in ImageJ (v1.50i, National Institutes of Health, Bethesda, USA). The CT values were averaged over a sphere of 2 cm diameter, centered in the middle of the insert. As the calibration curves in Fig. 2 show, the energy dependence vanishes when the DD reconstruction is applied. The right panel displays this implemented calibration given in eq. (1).

2.C | Study setup
All phantoms displayed in Fig. 3 were scanned with the SOMATOM Confidence CT. In general, images for 80, 100, 120 kVp, as well as 140 kVp were generated for the E30s (DD) kernel. The reference images (HU) were done with 120 kVp and the B40s kernel, as this represents the standard parameters in the clinical routine. The tube current was set between 140 and 300 mAs, depending on the thickness of the phantom, but held constant for each phantom. Slices of 3 mm thickness were acquired. All scans were imported into Eclipse (version 13.6, Varian Medical Systems, Palo Alto, CA, USA). The contouring was performed on the HU image and then copied to the DD images. The same procedure was followed creating the dose plans: Each plan was optimized and calculated on the 120 kVp HU image and transferred to all DD images. Then, each plan was recalculated holding all beam parameters constant.
The dose was calculated in the TPS Eclipse with the Anisotrop Analytical Algorithm (AAA) (v.13.6.23, Eclipse, Varian Medical Systems, Palo Alto, CA, USA), a 3D pencil beam superposition convolution algorithm. For this algorithm, only RED but not MD needs to be assigned to the CT voxels for a correct dose calculation. 9,10 Thus, the immediate relation between dose and the DD images which contain RED information is evaluated. The dose calculation grid size was chosen to be 0.125 cm.
To compare the different dose distributions, a dose difference between each plan on the DD reconstruction with the HU dose distribution as reference was created. Because the dose distributions were calculated on identically positioned phantom CT scans, which were performed without moving the phantom, the gamma criteria 11 which is typically used is not needed. However, passing rates for the dose differences are presented which equal a gamma analysis with 0 mm distance-to-agreement and 0.5% dose difference in the region with doses higher than 5% of the maximum dose.
Additionally, dose-volume histograms (DVH) were used to evaluate dose distributions in specific regions like the PTV and surrounding OARs as contoured in Fig. 3. For the general evaluation, dose profiles were used to further investigate the effect of high and low density materials on the dose distribution.

2.C.2 | Setup for metal implants
One of the main advantages of the implementation of the DD algorithm in the clinic is the reduction from many CT calibration curves to only one. Therefore, the effect of metal implants in combination with the DD reconstruction was tested, as the DD reconstruction could define the metal as high density bone due to the applied threshold. If this is the case, an inadequate interpretation of metal could wrongfully influence the reconstruction around these implants.
For this study, a third phantom was used. The CBCT Electron Den-  For a more anthropomorphic setup, the ART head was scanned with and without a brass insert simulating a dental prosthesis at 200 mAs. As differences due to the artifacts induced by metal ought to appear, an additional reconstruction with a metal reduction algorithm (Siemens Healthcare, Erlangen, Germany) was also evaluated.

2.C.3 | CARE dose option
To evaluate the possible dose savings when the DD algorithm is implemented, the tube-current as well as the voltage were varied.
The pelvis and thorax of the ART phantom was scanned again using CARE kV and CARE Dose4D (Siemens Healthcare, Erlangen, Germany     5) and for air +3.3% in the DD 140 (see Fig. 6). The dose difference in the depth dose curves for air is resolved beyond the air insert.

3.A.2 | Simple geometry
Whereas after the plaster insert, there is still a noticeable discrepancy in the solid water after the beam traverses the plaster insert.   lung as well as in the bladder, no deviation in the mean dose was found. The largest difference was observed in the trachea for the conformal thorax plan. This is due to two reasons: For one, one of the three beams aims directly at the trachea leading to a high dose in this region and additionally, the trachea is filled with air. With the results of the IMRT air cavity in mind, a higher dose has to be expected and is observed. Nevertheless, the values in Table 2 are in an order to be negligible.  However, a calibration for high density materials could be added, diverging from the straight line from eq. (1). If this is implemented,

3.B | Effects of metal implants
CTs with high density materials should only be acquired with one chosen tube voltage as the metal values are not independent of the tube voltage anymore as visible in Table 3 and Fig. 10.

3.C | Dose reduction with CARE dose
The use of the CARE kV and CARE Dose4D options allow the CT to acquire a high quality image with a potentially reduced dose exposure by choosing the kV and adapting the mAs to the given situation. While CARE Dose4D can be used with the conventional HU calibration, CARE kV can only be enabled in radiation therapy when the DD algorithm is implemented.
A comparison between the CT parameters estimated during the acquisition with and without the CARE options is given in Table 4. It shows a dose reduction via the reduction in the dose length product The passing rate for the dose difference between the dose calculated on these scans and the HU dose are in the same order of magnitude as all other DD scans. The image quality (see Fig. 11) was found to be sufficient.

| DISCUSSION
The Siemens DirectDensity TM algorithm was implemented and validated. With a simple phantom, the consistency of the RED conversion for the conventional HU to RED conversion and the Direct Density to RED conversion was tested. The results show small discrepancies that indicate that there will be discrepancies in the dose distributions. These were evaluated for different radiation modalities and setups as well as simple and also anthropomorphic geometries.
A direct correlation was observed between an overestimated RED and an overestimation in dose and vice versa. Air was found to produce the biggest dose discrepancies.  13 where the passing rates for the gamma criteria of 1 mm and 1% were found to be greater than 99%. The results presented in this work suggest that the gamma criteria tool to evaluate the effects of the direct density dose gives a good indication but for a complete analysis it is not sufficient.
The differences for the mean dose (see Table 3) are in agreement with results from Ref. 7  However, the intercept differs by 0.006 in the maximum case, as Ritter and Van der Heyden used 1 and we fitted 0.994 as intercept.
Chancing the intercept would change our results not significantly.
The deviations in air would decrease while the ones for bone would slightly increase in case of low tube voltages. Thus, CTs of other institutions with DirectDensity TM could be used with differences of less than 1% in the mean dose. If acceptable, this would benefit the patient by canceling the additional dose of another planning CT.
The second major benefit is the patient dose reduction by the implementation of DD with CARE kV and CARE Dose4D. This can reduce the dose significantly without reducing the image quality, allowing a more patient specific CT acquisition.
T A B L E 4 Comparison of the CT parameters before and after the CT options CARE kV and CARE Dose4D parameter optimization. A problem that could occur with the implementation are the danger of CT scans with an unsuitable tube voltage, resulting in an unusable scan and thus a rescan with additional dose to the patient.

| CONCLUSION
A CT reconstruction method for a direct translation between image values and relative electron density was implemented and found to provide sufficient image quality and dose calculation accuracy. The

CONF LICT OF I NTERESTS
We have no conflict of interest to declare.