PDF
Citation
Print
Introduction
Magnetic resonance image-guided radiotherapy (MRgRT) systems are becoming more important in radiation oncology [1-4]. It provided real-time soft tissue imaging via an onboard magnetic resonance (MR) scanner, which has better soft tissue visualization than conventional computed tomography (CT) or cone-beam CT imaging [5,6]. The superior contrast of tissue allows for the daily management of interfraction anatomical variation, particularly in large areas like the abdomen [3]. These imaging characteristics allow online adaptive radiation therapy (ART) to account for ongoing patient anatomy changes throughout the treatment, resulting in more accurate tumor targeting [1].
One of the most significant barriers to online ART clinical practice is the cost of increased time for recontouring, reoptimization, and dose calculation. The process of updating the electron density (ED) map is especially important [2]. The ED map is assigned daily based on a deformable registration of simulation CT and daily MR images [7]. Although the deformable registration algorithm is conducted, it is challenging to accurately deform even the air pocket in the bowl, which can cause significant changes in the ED map. Therefore, in the ART process, the air cavities of simulation CT and daily MR are manually contoured and bulk-assigned water and air density, respectively.
Recently, a synthetic CT image generation which is the direct conversion of MR into CT images based on a deep learning technique has been suggested by significant researchers in the radiotherapy field [8-12]. This approach has the potential to improve the dose calculation accuracy than the previous synthetic CT approach which delineated volumes with a homogeneous ED, referred to as bulk density [13]. To date, there have been few studies that have compared the influence of the ED overridden with the synthetic CT to the manually assigned method, particularly for the air cavity variation.
In this study, we retrospectively assessed the dosimetric impacts of ED maps derived from manually assigned volume and synthetic CT images on the position variability of the air cavity in pancreatic cancer MRgRT.
Materials and Methods
1. Patient selection and planning
Nine patients with pancreatic tumors treated for stereotactic MRgRT at Seoul National University Hospital were retrospectively registered after obtaining Institutional Review Board of the Seoul National University Hospital approval (IRB No. 1708-051-876). Informed consent was not required in this study because only image data were used and they presented minimal risk of harm to subjects. The patients initially had a CT and MR simulation. During the CT simulation step, the patients were positioned with MR dummy coils placed on the abdomen to imitate the MR simulation. The planning CT image acquisition was conducted with the Brilliance CT Big Bore (Philips, Amsterdam, Netherlands) scanner with 120 kVp, 2 mm slice thickness, and YA reconstruction kernel with iDose 4 level. Simulation MR images were obtained with the MRIdian (ViewRay Inc, Mountain View, CA, USA) Cobalt-60 system using the identical patient position and immobilization devices like the CT simulation.
The primary images for contouring targets and organs at risk (OARs) were chosen with simulation MR images while the CT images were deformable and registered to the primary images for defining ED. The gross tumor volume (GTV) was delineated using both CT and MR images, and the planning target volume (PTV) was established as a 4 mm expansion from the GTV. The prescription dose of 45–50 Gy, with a daily dose of 9–10 Gy, was delivered to the PTV using a 5-fraction regimen. The plan optimization and dose calculation were computed using the MRIdian system with a 0.35 T magnetic field condition. The plan was normalized to cover 95% of the PTV with 100% of the prescribed dose.
2. Electron density modification
We generated the additional dose distributions utilizing identical beam parameters except for the ED map. The first dose distribution was evaluated using CT-derived ED. It is known as “Original.” The second dose distribution is known as “bulk density assign, BDA.” The air cavities in both CT and MR images were segmented and manually adapted using image threshold. The air cavities in the CT images were virtually filled with water ED, whereas the cavities in the MR images were assigned to air ED. The third dose distribution was calculated using synthetic CT-driven ED, known as “sCT.” The synthetic CT images were created using a deep learning model trained with the CycleGAN architecture from simulation MR images. The structural framework of the deep learning network and training procedures were discussed in the previous study [12]. To modify the ED map, we modified the simulation CT images to synthetic CT images. Fig. 1 depicts the overall procedure of the study.
3. Dosimetry evaluation
We contrasted the dose-volumetric parameters of each plan to determine the effect of air cavity variation on planning. For target volume, the following parameters were calculated: near dose maximum (D5%) near dose minimum (D95%), and dose maximum. The target volume’s homogeneity index (HI) and conformity index were also computed, with the definitions as follows:
where D5 and D95 are the minimum doses in 5% and 95% of the target volume, respectively, and Dp is the prescribed dose [14]. The ideal value is zero when D5 and D95 are equal.
where VRI is the volume of the reference isodose and TV is the target volume [15].
Dose-volume histograms (DVHs) were estimated for the OARs and clinically significant parameters were extracted. The DVH parameters were statistically analyzed using the paired t-test and IBM SPSS Statistics v25 software (IBM Corporation, Armonk, NY, USA). Statistical significance was defined as P-values less than 0.05.
We also evaluated the DVH parameter difference on the residual values between the plans using an outlier detection algorithm. A Z-score method depends on the mean and standard deviation of the data to calculate central tendency. However, the drawback of the Z-score approach prevalently never detects an outlier if the dataset contains fewer than 12 samples. A modified Z-score approach was created to quantify in small-size datasets or data that is not typically distributed using median absolute deviation and median values rather than mean and standard deviation [16]. We used the modified Z-score method with threshold 2 to symbolize the DVH parameter discrepancy cases.
Finally, we conducted a 3D gamma analysis to create a map denoting the location of the points with a large dose difference using the PyMedPhys library (https://docs.pymedphys.com/). The gamma criteria of 1%/1 mm and 2%/2 mm with a 10% threshold were used.
Results
Fig. 2 depicts an example of a simulation MR (left), simulation CT (middle), and synthetic CT (right) slices from the OAR location. The red circles indicate a difference in the locations of air cavities between the simulation CT and MR images. However, the air region patterns were similar in the synthetic CT generated from the MR images.
Table 1 shows dose-histogram parameters that describe PTV dose coverage for the three plans. D95% averages were 48.80 Gy, 48.50 Gy, and 48.23 Gy for Original, BDA, and sCT dose distributions, respectively. The greatest deviation was observed in one patient, whose D95% to sCT was 1.84 Gy higher than in the Original plan. The average dose differences of Dmax to the Original plan according to ED modification were discovered to be less than 0.69 Gy. However, the DVH parameters before and after ED modifications were not observed to be statistically significant except for the HI.
Table 2 summarizes the dosimetric results of the ED modification for OARs. For the liver, the V25Gy values were 11.74 and 11.72 cc for BDA and sCT methods, respectively. There were no statistically significant differences in the dosimetric parameters between the Original and the two ED modifications.
Table 3 displays the modified Z-score results for each case based on the dosimetric differences from the Original. We boldly highlighted the outlier cases that exceed the threshold value (2.0). For case #4, two dosimetric parameters (D98% and D95%) are considered outliers at both BDA and sCT dose maps compared with the Original. Fig. 3a displays the DVH of the outlier case in the target volume. DVH curves are illustrated in Fig. 3b by selecting a case with a lower modified Z-score value for visual comparison.
Table 4 shows the average gamma analysis results for three dose distributions. Overall, for BDA and sCT, gamma analysis pass rates at 1%/1 mm between Original and ED modification plans were 96.86% and 96.38%, respectively. The lowest gamma analysis result at 1%/1 mm between Original and sCT plans was 92.22%, which increased to 97.84% when two ED modification methods were compared. For the remaining patients, the passing rates between BDA and sCT were also slightly higher (98.91%) than the Original plan. Generally, gamma analysis at 2%/2 mm was insufficiently sensitive to represent ED modifications (pass rates ranging from 98.90% to 99.99%).
Discussion
Because of its superior soft tissue contrast, the onboard MR imager improves the visibility of the tumor volume and related OARs during treatment. By providing volumetric MR images, these characteristics enable online adaptive radiotherapy for interfraction modification and intrafraction motion management [17]. Image alignment between simulation and daily MR images, followed by couch positioning, could enhance patient positioning. If the couch positioning is not compensating for OAR variation, reidentification of the target and OARs contours would be conducted and the dose distribution would be reoptimized. However, one major impediment to these processes in adaptive radiotherapy is the extensive workload required for contour editing and delineation [2]. Particularly, because the position of the air pockets is randomly distributed, it is critical to manually define the contour in each treatment [18]. Therefore, we retrospectively evaluated the dosimetric effects on modified ED maps that were either manually assigned to the air cavity region or altered by synthetic CT images generated with a simple endeavor.
Despite the lack of statistically significant differences in the dosimetric parameters between the Original and ED override plans, modifications in the air cavity positions during simulation or treatment cause dose differences in some cases. For example, the parameters are substantially impacted when the cavity region varies around the target volume. Those, on the other hand, have less influence because the changing position is further away from the region of interest. We addressed these cases using the modified Z-score approach to uncover the dosimetric parameter discrepancy from the Original. Fig. 4 depicts a case, designated as an outlier by the algorithm, in which the cavity locations differed between CT and MR simulation. In the synthetic CT image, it was validated that the position of the air was comparable to that of the MR. For this case, reassigning the air cavity regions or changing the synthetic CT images to modify the ED should be conducted to evaluate the dose evaluation correctly.
The MRgRT process with online adaptation could enhance treatment outcomes while reducing OAR toxicity. Michalet et al. reported the dosimetric advantage of MRgRT for pancreatic cancers [19]. The daily adaptation indicated the benefit in terms of tumor coverage and OAR savings. An important step in the online adaptive workflow is the redelineation of the target, OARs even air cavities. To lower this labor-intensive step, only a region enclosing the target volume by 2 cm [20] or 3 cm [21] is typically assessed for contour adaptation. This approach, however, may result in an inaccurate assessment of the dosimetric parameters for OARs located far from the target.
Recently, deep learning-based image generation demonstrated exceptional performance in the medical imaging field. The deep learning-based model’s major benefit is that it can directly convert MR images into synthetic CT images within seconds. With a large dataset, the model learned the complex nonlinear relationship between two modality images and generated the corresponding CT HU values from the MR image in near real-time. This technique eliminates dose calculation errors caused by image registration between planning CT and MR images. This could reduce the uncertainty caused by changes in organ location, particularly the unpredictable air cavities between simulation and treatment time. Our study demonstrates the dosimetric similarity with gamma evaluation between the ED modification by manually assigning the regions and the automatically generated synthetic CT images. Though this sCT technique has the potential to overcome the limitations, clinical implementation will be hampered by the integration of the sCT algorithm into the treatment machine.
The small number of datasets used for the dosimetric evaluation is one of the study’s limitations. The sample size was insufficient to generate a statistically significant difference between the original dose distribution and the two modified dose distributions for ED map modifications. However, we proposed an outlier detection strategy to determine which cases were significantly affected by the air cavity changes. Second, we compared the two ED modification methods using the simulation CT and MR images. These images were obtained at roughly one-hour intervals, which may have had less of an impact on the variability of the organ position. Dosimetric parameters such as air cavity variation may have a greater impact on daily adaptation planning. In the future, a comparative analysis will be performed using the adaptive plans and daily MR images.
XML Download