Towards a generalised development of synthetic CT images and assessment of their dosimetric accuracy

Abstract

Background: The interest in generating “synthetic computed tomography (CT) images” from magnetic resonance (MR) images has been increasing over the past years due to advances in MR guidance for radiotherapy. A variety of methods for synthetic CT creation have been developed, from simple bulk density assignment to complex machine learning algorithms.

Material and methods: In this study, we present a general method to determine simplistic synthetic CTs and evaluate them according to their dosimetric accuracy. It separates the requirements on the MR image and the associated calculation effort to generate a synthetic CT. To evaluate the significance of the dosimetric accuracy under realistic conditions, clinically common uncertainties including position shifts and Hounsfield lookup table (HLUT) errors were simulated. To illustrate our approach, we first translated CT images from a test set of six pelvic cancer patients to relative electron density (ED) via a clinical HLUT. For each patient, seven simplified ED images (simED) were generated at different levels of complexity, ranging from one to four tissue classes. Then, dose distributions optimised on the reference ED image and the simEDs were compared to each other in terms of gamma pass rates (2 mm/2% criteria) and dose volume metrics.

Results: For our test set, best results were obtained for simEDs with four tissue classes representing fat, soft tissue, air, and bone. For this simED, gamma pass rates of 99.95% (range: 99.72–100%) were achieved. The decrease in accuracy from ED simplification was smaller in this case than the influence of the uncertainty scenarios on the reference image, both for gamma pass rates and dose volume metrics.

Conclusions: The presented workflow helps to determine the required complexity of synthetic CTs with respect to their dosimetric accuracy. The investigated cases showed potential simplifications, based on which the synthetic CT generation could be faster and more reproducible.

Introduction

Image guidance with magnetic resonance (MR) for radiotherapy is currently one of the most promising options for improvement in treatment outcome. However, the MR images acquired before or during the treatment cannot directly be used for dose planning, since to this end, electron density (ED) is the dominant quantity needed. The latter is conventionally derived from a computed tomography (CT) image via a look-up table. However, for situations where dose calculation on MR images is intended or cannot be avoided, the lack of ED information can be overcome by MR-based, so-called “synthetic CTs” (sCTs) (sometimes also called “pseudo CT” or “substitute CT”).

A number of approaches for the computation of sCTs have been suggested and investigated in the literature over the last years. They can be divided into atlas-based approaches [1–3] and machine learning approaches [4–9]. The latter range from simple bulk density assignment (BDA) [4,5] over classification and regression [6–8,10] to complex neural networks [9,11]. A recent review [12] summarises the different methods and their specified accuracy. For photon treatments, dose differences <2% can be observed for most cases. Commercial products are now also available on hybrid machines using BDA, as well as for pre-treatment MR imaging off-site (hybrid method [13], BDA [14]). For the latter, dosimetric evaluations were recently published by Tyagi et al. [15] and Persson et al. [16], demonstrating the applicability of the presented techniques.

The usage of BDA specifically for pelvic cancer has been widely investigated in the literature, however mostly for water equivalent (WE) bulk densities [4,17] or WE and bone [4,5,18] as surrogate EDs. The commercial approach described in Köhler et al. [14] increased the complexity of the sCT by using five different tissue classes for BDA. The next increase in complexity is the dual model [19,20], which separates bone (polynomial fit) and soft tissue (interpolation between fat, muscle and urine) for sCT generation. Machine learning and image registration are the most complex and computational demanding approaches. However, at some point an increase in complexity of the sCT generation might not result in a relevant increase in dosimetric accuracy any more. This trade-off has not yet been analysed systematically.

In this study, we therefore investigate the necessary level of complexity of sCT generation for clinical purposes. The generation of sCTs includes, independently from the applied approach, two aspects: (1) the complexity of ED information needed for a specific treatment, i.e., number of tissue classes, and (2) the uncertainties in relating MR contrast and CT numbers. While these two aspects are usually treated simultaneously, we present a methodology to decouple the two factors, which facilitates the isolated investigation of the required granularity in ED information. Through the decoupling, our study may serve as a fundamental guideline for the design of MR to sCT conversion algorithms. Through the consideration of clinically meaningful treatment uncertainties, it may inform future discussion about the practically meaningful complexity of sCTs in clinical MR-only workflows. These represent an upper limit for the accuracy of the sCTs.

Material and methods

We created a set of seven simplified CT images (simCTs) each from the original CTs of six pelvic cancer patients with different numbers of tissue classes. The definition of the tissue classes originated from the structure of the relative ED (relative to water, rED) histogram shown exemplarily in Figure 1 for the pelvis: two tissue types (fat and non-fat soft tissue) dominate the histogram in the form of two Gaussian-shaped peaks. This simple and reproducible shape suggested that only few rED classes are sufficient, representing either both peaks or a ratio of them in only two or one class, respectively. In the context of MR-based treatment planning, these classes can correspondingly be identified in conventional T2-weighted MR contrasts, as the 2 D histogram in Figure 1 suggests. T2-weighted sequences or fat/water separating sequences are especially useful to separate tissue similarly to the CT-based fat/soft tissue separation. Air and bone are well differentiable in the rED, but not in the MR signal, resulting in a smeared, low signal in Figure 1 around the main peaks. As also pointed out by others [6,7,12,22], the main challenge of MR-based treatment planning is thus the differentiation of air and bone and their assignment of rED. To assess the importance of air and bone classes in the dose calculation, a simCT with four classes including air and bone was created as well.

Figure 1. Histogram of relative electron density derived from the CT of the pelvis via a clinical HLUT (bottom right) and histogram of MR signal intensity from the corresponding T2-weighted MR image after bias field correction (top left). The top right image shows the 2 D histogram of the relative electron density and the MR signal. The MR image was acquired in the same position as the CT and close in time to the CT acquisition, as described in detail in Bostel et al. [21].

Patient selection

Six patients who underwent intensity-modulated radiotherapy (IMRT) of pelvic tumours at the German Cancer Research Center (DKFZ), Heidelberg, were selected for this study. Indications and fractionation scheme information can be found in Supplementary Table 1. Patients were part of a MR guidance study [21] and gave informed consent for the retrospective usage of their image data. Planned clinical dose distribution and contours for the target and organs at risk (OARs) were available from experienced radiation oncologists created within the clinical treatment planning system (TPS) RayStation (RaySearch Laboratories AB, Stockholm, Sweden). More details on patient selection, imaging and treatment planning can be found in [21] and [23]. The treatment plan optimisation and dose calculation was performed with the open-source treatment planning system matRad [24] (https://www.matRad.org), which offers higher flexibility than commercial TPS, for example access to beam weights and beam weight initialisation for dose optimisation. The simulation of uncertainties and subsequent data analysis was performed in Matlab (The Mathworks, Inc., Natick, USA). The matRad program is based on Matlab, therefore the transition of dose calculation results to analysis was straight forward. Using a clinical Hounsfield lookup table (HLUT) from DKFZ, all original CTs were converted into rED ($\widehat{\mathrm{\rho }}$) images, which are the basis for all simplified rED images (simED).

Bulk density determination

For each patient, the histogram of rED inside the body contour was analysed. The two dominant peaks representing fat and remaining soft tissue, which includes organs, muscle, etc., were fitted with a sum of two Gaussian functions to find the peak positions $\widehat{\mathrm{\rho }}$_{fat_pat} and $\widehat{\mathrm{\rho }}$_{soft_pat}. Additionally, the crossing point of the two Gaussians was determined as the separation point between fat and remaining soft tissue ($\widehat{\mathrm{\rho }}$_{sep_pat}). The peak and minimum positions were also averaged over the six patients ($\widehat{\mathrm{\rho }}$_{sep_coh} $\widehat{\mathrm{\rho }}$_{fat_coh}, and $\widehat{\mathrm{\rho }}$_{soft_coh}).

For the creation of simEDs, the body in the original rED image was segmented using histogram-based thresholding into four compartments: air ($\widehat{\mathrm{\rho }}$ < 0.6), fat (0.6 ≤ $\widehat{\mathrm{\rho }}$ < $\widehat{\mathrm{\rho }}$_{sep_pat}), soft tissue ($\widehat{\mathrm{\rho }}$_{sep_pat}≤ $\widehat{\mathrm{\rho }}$ ≤ 1.07) and bone ($\widehat{\mathrm{\rho }}$ > 1.07). Based on these, seven different simEDs were created by overwriting the compartments with bulk density values. A summary over all simEDs, their compartments and respective rED values can be found in Table 1. The four main steps of complexity represented by the simEDs are water-equivalent (WE), single-tissue (1t), two soft tissues (2t), and two soft tissues plus air and simplified bone (4t). In view of MR-based treatment planning, WE, 1t and 2t could be created just by classification on common MR sequences with good fat/water separation. For simED_4t, simple segmentation on the MR image is not directly feasible due to comparably low signal intensity in both air and bone. Either additional MR-sequences (e.g., UTE or ZTE) or other bone/air separation techniques have to be employed, such as manual bone segmentation [4,25,26].

Table 1. Overview of the relative electron density ($\widehat{\mathrm{\rho }}$) values assigned to the four compartments of the different simEDs.

Cohort-specific	WE	1t_coh	2t_coh	4t_coh
Air	1	$\widehat{\mathrm{\rho }}$sep_coh	$\widehat{\mathrm{\rho }}$fat_coh	0.04
Fat				$\widehat{\mathrm{\rho }}$fat_coh
Soft			$\widehat{\mathrm{\rho }}$soft_coh	$\widehat{\mathrm{\rho }}$soft_coh
Bone				1.25
Patient-specific		1t_pat	2t_pat	4t_pat
Air	-	$\widehat{\mathrm{\rho }}$sep_pat	$\widehat{\mathrm{\rho }}$fat_pat	0.04
Fat				$\widehat{\mathrm{\rho }}$fat_pat
Soft			$\widehat{\mathrm{\rho }}$soft_pat	$\widehat{\mathrm{\rho }}$soft_pat
Bone				1.25

Optimisation

The treatment plan for original rED images was optimised in matRad to reproduce the dose distribution from the clinical treatment plan, considering dose constraints from QUANTEC [27] as well (from now on referred to as reference dose distribution).

Independent treatment plan optimisation was performed on all simEDs using the identical dosimetric treatment plan objectives and constraints. The beam weights from the original plan were used as initial beam weights for simED optimisation to accelerate consistent convergence. The gamma pass rates (GPRs, 2 mm/2%-criterion) inside the body contour and the PTV comparing the dose distribution from the original rED and the simEDs were above 99%, which indicates treatment plan consistency. As a second step, the associated optimal fluences from each simED were used to recompute the dose distribution on the original rED image (from now on referred to as the dose from simED) and compared to the reference dose distribution. Finally, the dosimetric differences between the simEDs and the original rED image were evaluated.

Uncertainty scenarios

The simulation of uncertainties is key for meaningful interpretation of the clinical relevance of the observed discrepancies between simEDs and original rED. Two types of uncertainties were simulated for comparison to simEDs: geometrical setup and range uncertainties. Details how the simulation was performed can be found in the Supplementary material. The uncertainties were simulated both for the original rED image and the simED images.

Analysis

To analyse the resulting dose simulations, 3 D gamma tests [28] using 2 mm distance and 2% global dose criteria with a 2% lower threshold, and corresponding GPRs were calculated within the PTV volume. SimEDs and the uncertainty scenarios from both the original rED image and the simEDs were compared to the original rED each. The latter resulted in GPR distributions, which are visualised as boxplots.

Further, dose volume metrics D_x (the dose values that x% of the volume receives) were calculated for the PTV and the OARs from dose from the original rED image, the simEDs and the uncertainty scenarios. For target volumes D₂, D₅₀, and D₉₈, for OARs D₂ and D_mean were used, following the ICRU Report 83 [29].

Again, the distribution of D_x values from uncertainty scenarios simulating positioning errors are visualised as boxplots, whereas the uncertainty scenarios simulating HLUT shifts are shown as errorbars. The HLUT shift range and the interquartile range (IQR) of position shifts (black box in the boxplots) from the original rED image defined two confidence intervals, which are used as acceptance criteria for the simEDs, i.e., the criterion is passed if the D_x values lie within the IQR of position shifts or the HLUT shift range. In Figure 3, the two confidence intervals are included as shaded areas. However, some intervals have a “glass top” or “glass bottom” indicated by the black arrows in the figure. This means that also values above or below the interval are accepted, if this conforms or reduces dose to the target or OAR, respectively. Accepting an actual under- or overdosage is prevented by the usage of the combination or two or three D_x values per simED and volume.

Results

Bulk density determination

The detailed results from the histogram fits are given in the Supplementary Table 2. The rED of the fat peak fit position ($\widehat{\mathrm{\rho }}$_{fat_pat}) varies between 0.924 and 0.948, with a mean ($\widehat{\mathrm{\rho }}$_{fat_coh}) of 0.934 ± 0.008. The soft tissue fit peak ($\widehat{\mathrm{\rho }}$_{soft_pat}) ranges from 1.019 to 1.034, with a mean ($\widehat{\mathrm{\rho }}$_{soft_coh}) of 1.027 ± 0.005. The minimum $\widehat{\mathrm{\rho }}$_{sep_pat} is smaller than water ($\widehat{\mathrm{\rho }}$ = 1) for all six patients, ranging from 0.970 to 0.980 with a mean ($\widehat{\mathrm{\rho }}$_{sep_coh}) of 0.973 ± 0.004.

Gamma analysis

An example for the gamma images is given in Figure 1 in the Supplementary material for one slice from patient 1 for simED_{1t_coh}. The gamma index failure can be linked to the anatomy, as can be observed in the figure when comparing the simED and the corresponding original rED slice. Failures occur mainly downstream of a large air bubble (white arrows) and where the beam travels significantly through cortical bone (yellow arrows), which mostly occurs when the beam is parallel to the surface of the bone.

Figure 2 (left) shows the comparison between the GPRs of the simED (symbols) and the distribution of GPRs under the influence of positioning uncertainties (boxplots) for the original rED image for the PTV. For clarity, the uncertainty scenarios of the individual simEDs are not included here and shown exemplarily for patient 1 on the right. Uncertainties from the HLUT for the original rED image are also not shown here, as they are above 99.98% in all cases.

Figure 2. Left: Gamma pass rates (GPRs) of the PTV for the simEDs compared to the original rED image are shown as symbols. GPRs from sampled shifts of the original rED image are shown as boxplots. Uncertainty sampling is shown exemplarily for patient 1 on the right: positioning sampling is included as boxplots, HLUT sampling as errors bars.

The GPRs of simED_4t are almost 100% for all patients. Worst results are obtained by both patient and cohort specific simED_1t, however simED_WE has higher GPRs than simEDs_1t in all cases. All simED_2t show GPRs above 96%. The boxplots of the GPRs obtained by the positioning uncertainty sampling visualise that the GPR drops down to ∼70% in extreme cases for the PTV. The IQR can reach down to ∼85% for individual patients as well, which is worse than the GPRs observed for most simEDs.

On the right, one example is given for GPRs of simEDs compared to the original rED image including the uncertainty sampling for patient 1. For simEDs with lower GPRs, the uncertainty scenarios introduce even larger uncertainties, as for example for simED_1t. However, for high GPRs of the simED, such as 2t and 4t, the sampled shifts have similar influence as the sampled shifts of the original image and the HLUT shift only has a very small influence.

Dose volume metrics

Table 2 summarises the results of the mean relative dose difference of specific D_x for the PTV and three OARs of the six patients. These OARs are available for all selected patient cases independent from indication and are thus used for evaluation. Patient 4 does not receive any dose to the spinal cord, thus the values for the spinal cord are averaged only over the remaining five patients.

Table 2. Mean relative dose differences for a set of dose volume metrics (DVMs) averaged over the patient cohort.

		Mean relative dose differences [%]
Volume	DVM	1t_coh	2t_coh	4t_coh	1t_pat	2t_pat	4t_pat	WE
PTV	D2	−0.7	−0.1	0.3	−0.7	−0.1	0.2	0.4
	D50	−1.5	−0.6	0.3	−1.5	−0.6	0.2	−0.3
	D98	−1.4	−0.4	0.3	−1.4	−0.4	0.3	−0.3
Bowel	D2	−0.3	−1.8	0.3	−0.3	−1.6	−1.5	0.8
	Dmean	−0.3	−1.1	0.4	−0.4	−1.0	−1.0	0.9
Bladder	D2	−2.2	−1.0	0.3	−2.2	−1.1	0.2	−1.0
	Dmean	−1.8	−2.2	0.3	−1.8	−2.1	−1.3	−0.6
Spinal cord	D2	−2.3	−1.1	−0.5	−2.3	−1.1	1.3	−0.2
	Dmean	−1.8	−0.3	−0.2	−1.8	−0.3	1.7	0.0

Largest deviations from the reference dose distribution for the PTV can be observed for simED_1t, the smallest for simED_{4t_pat}. The deviations for simED_2t, simED_4t and simED_WE are, however, smaller than 1%. Among the OARs, mostly negative dose differences can be observed. The largest positive deviation can be observed for simED_{4t_pat} for the spinal cord.

An example for the D_x values including the simulated uncertainties is shown in Figure 3 for patient 2. Here, the trends of the D_x deviations from the reference dose distribution that have been presented in table 2 can be recognised. The width of the envelopes for D₂ and D₅₀ for the PTV are dominated by the HLUT sampling, whereas for D₉₈, the position sampling creates a wider span of D_x values. All D₂ and D₉₈ for the PTV lie within the envelopes and the “glass bottom/top”, whereas for D₅₀, the values of the DVM for simED_1t lie below their respective envelopes. Similarly, an example for an OAR is given below for the bladder of the same patient. Here, no exceeding of the dose envelopes and the glass bottoms can be observed.

Figure 3. D₂, D₅₀ and D₉₈ of the PTV (top) and D₂ and D_mean of the bladder (bottom) of patient 2 for all simEDs. The uncertainty sampling is included as boxplots for the positioning sampling, and as error bars for the HLUT sampling. The shaded regions and the black arrows represent the acceptance envelopes based on the uncertainty sampling of the original ED image (more details in the Supplementary material).

Generally, the PTV criterion was the most critical one, resulting in most criterion failures. The two simED_4t always passed the PTV criterion, simED_WE and simED_2t performed almost as good with one or two failures each. Only simED_1t showed negative results for the PTV for all tested cases. Regarding the OARs, simED_{4t_pat} showed best results with no failures at all, followed by simED_{4t_coh} with one failure for the bowel for patient 1. While no criterion failures were observed for the bladder for all simEDs, two failures were observed for the spinal cord (one for simED_{2t_coh} and one for simED_WE) and one to three failures for the bowel for all simED except simED_4t.

Discussion

We illustrated the described workflow using a dataset of six pelvic patients. The pelvis is a challenging part of the body, as it includes potentially large variations. The tissue composition in the beam’s path can vary strongly depending on the patient’s weight, bladder filling or bowel gas. Such anatomical differences were observed between the six patients, the bladder volume for example varied between 84 cm³ and 397 cm³. Nevertheless, the results for each simED are consistent among the patients. This suggests generalizability of the results and the gain in informative value from much larger patient cohorts can be expected small.

The ED classification in this study was based on the frequency distribution of rEDs for the pelvic region. We assumed that only the prominent peaks for fat and non-fat soft tissue are of major importance for the planning process. Simplified air and bone models were also included as simED_4t, to incorporate the strong radiological impact of those tissues despite their small volume ratio. Consequently, this very aspect of our methodology does not represent a general solution for the whole body, as it can be expected that some tissues, for example lung, are not included in the two peaks. However, the general approach facilitated in our study here, i.e., finding an optimal number of tissue classes by comparison to other sources of uncertainty, could also be transferred to other sites.

Bulk density determination

Cohort mean values for BDA are advantageous when it comes to sCT creation from MR images where no patient specific ED values are available any more. Their applicability, especially for fat and soft tissue, was motivated by the small inter-patient variation in the peak positions from the histogram analysis (maximum 0.9% relative deviation). The values for $\widehat{\mathrm{\rho }}$_fat were slightly smaller than the respective values derived from the ICRU report 46 [30], as a consequence of the applied HLUT. This is, however, not critical to study, since the same HLUT was used consistently for all data and optimisations. The investigation with patient specific bulk rED values served as an intermediate reference, being closer related to the reference rED image via the histogram than the cohort mean rED values. It should be mentioned that $\widehat{\mathrm{\rho }}$_sep (also used as a BD for simED_1t) was introduced as the best approximation to separate fat and remaining soft tissue, but only partly reflects the volume ratio between fat and soft tissue, as for example a mean value would do. SimED_WE was found to perform better than simED_1t in GPR analysis and D_x analysis. We believe that one reason for this is the overestimation of the fat volume ratio in the beam’s path compared to the whole body, which is reflected by $\widehat{\mathrm{\rho }}$_sep being smaller than 1. In the beam’s path, the fraction of traversed fat is less than traversed non-fat soft tissue, therefore the value for the bulk rED is underestimated.

Analysis

The qualitative analysis of the gamma images revealed the sources and regions of GPR failure for the different simEDs compared to the original rED image. Apart from the simED_4t, all gamma images show failure in vicinity of air bubbles (e.g., in the lower gastrointestinal tract) and after considerable path lengths in cortical bone. The latter can be assumed to be reliable for each simED compared to the original rED image, whereas failures from air are comparable to the influence of inter- and intra-intestinal movement. These types of failure occur during standard fractionated treatment and are accounted for in the PTV margin. Path length variations on the other hand are not explicitly accounted for in the PTV margin, and, although a deadening effect occurs for slight variations, strong variations are not covered due to their unpredictability and locality. To better separate accounted influences from true failure from simEDs, the analysis of GPRs of the CTV or GTV would be advantageous, however those contours were not available for all patients. Both two and four tissue simEDs show acceptable GPRs above 96% in all patients. Fewer tissue classes lead to smaller GPRs for some patients, which was considered unacceptable. For the PTV GPR analysis and comparison to sampled shifts, it can be observed that the simEDs have less impact on the GPR than most of the simulated uncertainty scenarios.

The mean relative dose differences in Table 2 showed that the impact of two or more tissue classes can be considered negligible, as the respective deviations for the PTV are smaller than 1%. For simED_4t, all deviations are even below 0.3%. Thus, simED_{4t_coh} can be recommended, as also the D_x values for the OARs are below 0.5% deviation. For simED_{4t_pat}, deviations up to 1.7% occur, however these are all covered by the acceptance criterion defined by the uncertainty sampling.

Most D_x criterion failures were observed for the bowel. However, that usage of the D_x criterion for the bowel is debatable. Clinically, a V_x criterion (i.e., the volume receiving x Gy) is used rather than a D_x criterion. This, however, was not applicable here due to incomplete imaging of the organ in the original data set. Criterion failures for the PTV occurred mostly for simple rED images with one tissue class, which is in agreement with previous results from the GPRs and the mean relative dose differences. Also, simED_4t showed slightly better results than simED_2t for the D_x analysis, similar for GPR analysis. A noticeable characteristic of the simEDs regarding the sampled shifts can be observed: the comparable sizes of simEDs and reference boxplots in all investigated cases indicate that positioning uncertainties can also be detected with consistent use of simEDs, i.e., comparing the simED in original position to shifted position, which is desirably detected with sCTs. The difference with respect to the original rED image will add to the sampled values by shifting the boxplot up- or downwards in their entirety. SimEDs consisting of two tissue classes yield acceptable deviations from plans optimised on a reference CT for the specific case of pelvic tumours according to the applied criteria. Both GPR and D_x analysis results were comparable between patient specific and cohort simEDs, supporting the hypothesis that general values for a population are sufficient for treatment planning of pelvic tumours with simED images. This can reduce the effort per patient to only thresholding to detect the corresponding compartments and facilitates the creation of simEDs from MR images later on, when no patient specific EDs from reference CT scans are available.

These results suggest that even with a few tissue classes patients would benefit from MR guidance, which increases the geometrical confidence during irradiation, and dose re-calculation and re-optimisation (if necessary) with simple rED images. A further decrease of dosimetric accuracy due to interplay effects of simple rED images and patient shifts, as shown in Figure 2 (right), is not to be expected when generating the simED from MR images. The motivation to use MR guidance is precisely its ability to detect and thus reduce uncertainties such as position shifts or anatomical changes (not included in this study). It should be noted here that the generation of a sCT from a MR image will also suffer from MR artefacts, such as distortion or signal non-uniformity. These artefacts will further influence the quality of the sCT, but they are not part of this study, since their correction can be completely separated from the conversion of MR intensity to rED, for example by technical improvements [31] or pre-processing of the MR [32].

The simED generation in this study also showed that mis-classification of single voxels is a minor problem for the resulting dose. The two Gaussian fit curves in the histogram always overlap, suggesting that voxels from both tissue classes are falsely segmented by the threshold. Also, the strong simplification of complex bone structures yielded sufficient accuracy for treatment planning according to our criteria. This plays into the hands of MR-based treatment planning, where a perfect segmentation of the compartments, especially bone, is even harder to achieve than for CT images, for example due to susceptibility artefacts.

Conclusion

We present an approach to systematically study the accuracy achieved by algorithms for synthetic CT generation. The simulation of uncertainties (HLUT and positioning) serves hereby as a reference comparison for a clinically meaningful interpretation of results. For the six patient cases presented here, simplified versions of CT or rED images yielded comparable dosimetric results to the original image and introduced smaller deviation from the reference dose than positioning and HLUT uncertainties. This indicates that simplified CT images with a minimum of four bulk densities (air, fat muscle, bone) are sufficient for MR-based treatment planning for intensity modulated radiotherapy of pelvic tumours.

Our approach to determine the number of tissue classes can seamlessly be adapted to other data sets to enhance the explanatory power, or to other sites of the body. It provides structured information about the requirements regarding MR sequences for MR-based treatment planning, and thereby enables the identification of a suited trade-off between model complexity for simCT generation and accuracy for MR-based planning in routine treatments.

Disclosure statement

No potential conflict of interest was reported by the authors.