Evaluation of CBCT scatter correction using deep convolutional neural networks for head and neck adaptive proton therapy

Abstract

Adaptive proton therapy (APT) is a promising approach for the treatment of head and neck cancers. One crucial element of APT is daily volumetric imaging of the patient in the treatment position. Such data can be acquired with cone-beam computed tomography (CBCT), although scatter artifacts make uncorrected CBCT images unsuitable for proton therapy dose calculation. The purpose of this work is to evaluate the performance of a U-shape deep convolutive neural network (U-Net) to perform projection-based scatter correction and enable fast and accurate dose calculation on CBCT images in the context of head and neck APT. CBCT projections are simulated for a cohort of 48 head and neck patients using a GPU accelerated Monte Carlo (MC) code. A U-Net is trained to reproduce MC projection-based scatter correction from raw projections. The accuracy of the scatter correction is experimentally evaluated using CT and CBCT images of an anthropomorphic head phantom. The potential of the method for head and neck APT is assessed by comparing proton therapy dose distributions calculated on scatter-free, uncorrected and scatter-corrected CBCT images. Finally, dose calculation accuracy is estimated in experimental patient images using a previously validated empirical scatter correction as reference. The mean and mean absolute HU differences between scatter-free and scatter-corrected images are −0.8 and 13.4 HU, compared to −28.6 and 69.6 HU for the uncorrected images. In the head phantom, the root-mean square difference of proton ranges calculated in the reference CT and corrected CBCT is 0.73 mm. The average 2%/2 mm gamma pass rate for proton therapy plans optimized in the scatter free images and re-calculated in the scatter-corrected ones is 98.89%. In experimental CBCT patient images, a 3%/3 mm passing rate of 98.72% is achieved between the proposed method and the reference one. All CBCT projection volume could be corrected in less than 5 seconds.

1. Introduction

The treatment of head and neck cancers with radiation therapy is often challenging due to the proximity of organs at risk (OAR) to target volumes. In such scenarios, intensity modulated proton therapy (IMPT) has been shown to be superior to intensity-modulated radio therapy (IMRT) or volumetric-modulated arc therapy (VMAT) in terms of OAR sparing and integral dose mitigation (Barten et al 2015, Leeman et al 2017). However, the high dose gradients created with IMPT make it sensitive to set-up variations and anatomical changes. The latter is especially important for head and neck patients, who are prone to weight loss and tumor shrinkage. Over the course of a fractionated treatment, these effects can severely reduce the benefits of IMPT over photon therapy (Gora et al 2015g, Szeto et al 2016).

In order to keep the dosimetric advantages of proton therapy for head and neck patients amid set-up variations and anatomical changes, adaptive proton therapy (APT) has been proposed (Simone II et al 2011, Veiga et al 2015, Albertini et al 2020). APT can be performed offline, with a complete re-planning performed when the original treatment does not meet the clinical objectives (Simone II et al 2011, van Kranen et al 2013), or online, at each fraction (Bernatowicz et al 2018, Botas et al 2018). Online APT has the advantage of neither interrupting nor delaying the treatment delivery and allows a plan to exactly match the patient set-up and anatomy on each day (Lim-Reinders et al 2017, Nenoff et al 2019). A recent study from our group demonstrated the feasibility of online APT for head and neck patients, significantly improving treatment quality in the presence of inter-fractional geometry changes (Botas et al 2018).

One crucial element for online APT is the access to daily volumetric images of the patient, in order to perform dose calculation and plan adaptation. Such data is readily available in some proton therapy centers using on-board cone-beam computed tomography (CBCT). CBCT images are currently solely used for patient alignment and anatomy monitoring, although their potential for dose calculation has been investigated in the past years (Park et al 2015, Landry et al 2015, Veiga et al 2015). The main obstacle preventing the direct use of CBCT images for dose calculation is the inaccuracy of the reconstructed Houndsfield Unit (HU), affected by several imaging artifacts introduced by effects such as x-ray scatter within the patient and the detector panel. These artifacts are a major limitation in using CBCT for APT, where proton range uncertainties are strongly related to the accuracy of the stopping power mapping within patients (Paganetti 2012).

Several methods have been proposed to enable proton therapy dose calculation on CBCT images. Veiga et al (2015), Landry et al (2015) and Kurz et al (2016) investigated deformable image registration (DIR) of the planning CT to the daily CBCT in order to create a so-called synthetic CT or virtual CT. This approach is however not suitable for cases where fundamental differences between the daily CBCT and the planning CT are observed, such as air pockets or substantial weight loss. Park et al (2015) investigated a prior-based scatter correction, using the virtual CT as a first estimation of the primary signal in CBCT projections. This method was shown to be suitable for proton therapy dose calculation, although being also sensitive to the accuracy of the DIR (Kim et al 2016). The use of Monte Carlo (MC) simulations to predict scatter distributions in CBCT projections has also been shown to provide accurate image correction (Mainegra-Hing and Kawrakow 2008, Jarry et al 2006), although this approach has not been considered for APT, as it would typically require a computation time of several hours. Indeed, online APT requires both fast and accurate scatter correction of CBCT images in order to be clinically relevant.

More recently, approaches making use of deep learning for CBCT image correction have been investigated (Kida et al 2018, Hansen et al 2018, Maier et al 2019, Liang et al 2019, Harms et al 2019, Kurz et al 2019) and some have been evaluated for proton therapy dose calculation. Deep learning approaches offer the advantages of not relying on a planning CT after training and rapid image correction in a fraction of the time required by analytical and MC based approaches. Hansen et al (2018) have investigated the use of a U-shape deep convolutional neural network (U-Net) (Ronneberger et al 2015) trained to reproduce the projection-based scatter corrections from Park et al (2015) in pelvis patients. Their method was shown to be suitable for VMAT dose calculation of prostate treatments, but not satisfactory for proton therapy. Landry et al (2019) later demonstrated that this same U-Net architecture performed better when trained on reconstructed CBCT images rather than projections. Similarly, Maier et al (2018), Maier et al (2019) evaluated the performance of a U-Net trained to reproduce MC estimations of scatter distribution from raw projections in real-time, although the performance of their approach for dose calculation was not evaluated. Thummerer et al (2020) compared the performance of an image-based U-Net to DIR and analytical image correction, obtaining best accuracy for the two first approaches, albeit no projection-based approach was included in this study. Finally, Harms et al (2019), Liang et al (2019) and Kurz et al (2019) evaluated the performance of generative adversarial networks (GAN) to convert CBCT images into synthetic CTs. The latter study evaluated the accuracy of dose calculation in the synthetic CT of the pelvis region, which was once again shown to be sufficient for VMAT, but not for proton therapy.

In this study, we evaluate the performance of a U-Net trained to reproduce MC-based CBCT scatter correction in the context of head and neck APT. The implemented network follows the one of Maier et al, using MC simulated CBCT projections to train the model. The robustness of the method is investigated by evaluating the impact of some key features of the network on HU accuracy. A comprehensive evaluation of the dosimetric accuracy for head and neck IMPT is performed using MC simulated scatter free CBCT as ground truth. Range accuracy is experimentally evaluated using CBCT and CT images of an anthropomorphic head phantom containing a human skull. Finally, the U-Net based scatter correction is compared to a reference prior-based method in patient CBCT images. To the best of our knowledge, this study is the first to evaluate the performance of a MC based U-Net scatter correction for dose calculation.

2. Materials and methods

2.1. Patient data

A total of 48 head and neck patients originally treated with VMAT at the Massachusetts General Hospital between February and December 2019 were considered in this retrospective study. CT images were acquired for planning purposes in a wide bore GE scanner using a 140 kVp spectrum. The images were reconstructed with a pixel size ranging from 0.6 to 0.98 mm and a 2.5 mm slice thickness. These image volumes served to define the patient geometries used to simulate CBCT projections with Monte Carlo. Patients were distributed in training (29), validation (9) and testing (10) sets, following the customary 60/20/20 partitioning scheme.

In addition to the MC simulated images, CBCT images of a subset of 3 representative cases from the test patients cohort were used for comparison with a prior-based scatter correction approach. These images were obtained on a Elekta XVI system using a 100 kVp tube voltage, 10 ms exposure and 10 mA tube current, 220 degree acquisition with centered panel position, 20 cm collimator and no bowtie filter. CBCT images were reconstructed from raw and scatter-corrected projections using the FDK algorithm implemented in the open source toolkit RTK (Reconstruction ToolKit) (Rit et al 2014), using a ramp filter with a Hamming window (cutoff frequency of 5.0) and a truncation correction following the work of Ohnesorge et al (2000) approximating the truncated anatomy within a distance equivalent to 10% of the projection size. A 5× 5 median filter was applied to the projections before reconstruction, similar to what is done by the XVI reconstruction algorithm.

2.2. Empirical projection-based scatter correction

The performance of the MC trained U-Net scatter correction is evaluated in patient data using a prior-based scatter correction as reference, since no absolute ground truth is available. This method was proposed by Niu et al (2010) and validated for proton therapy dose calculation by Park et al (2015). A brief description of this approach is performed here, but readers are encouraged to see the original publications for more details. This method was chosen over the deformation of the planning CT considering the important interfractional anatomical differences that can be observed in the head and neck region and to facilitate comparison with recent work that have used the same approach as ground truth Hansen et al (2018), Landry et al (2019)

A raw projection generated by a CBCT system is described by:

$$p_{\text{raw}}=-\ln \left(\frac{I_{\text{raw}}}{I_0}\right),$$ (1)

where I_raw is the total intensity measured by the detector and I₀ is the open field intensity. The raw intensity I_raw can be decomposed into a scatter component S and a primary, scatter free component I_prim:

$$I_{\text{raw}}=I_{\text{prim}}+S.$$ (2)

The purpose of projection-based scatter correction methods is to obtain an estimation of the scatter distribution $\widehat{S}$, which can be subtracted from I_raw in order to create a scatter-corrected intensity map I_corr:

$$I_{\text{corr}}=I_{\text{raw}}-\widehat{S},$$ (3)

from which a scatter corrected projection is created with:

$$p_{\text{corr}}=-\ln \left(\frac{I_{\text{corr}}}{I_0}\right).$$ (4)

The workflow used in the prior-based scatter correction is summarized in figure 1. First, the raw CBCT projections are reconstructed to create an uncorrected volume named CBCT_raw. The planning CT (pCT) is then deformed to CBCT_raw, to create a volume named virtual CT (vCT). Synthetic projections of the vCT volume are generated with digitally reconstructed radiographs (DRR) to create I_vCT, an a priori estimation of the scatter-free intensity I_prim. From there, the scatter distribution ${\widehat{S}}_{\text{prior}}$ is estimated as the low frequency deviations between I_raw and I_vCT:

$${\widehat{S}}_{\text{prior}}=f\left(I_{\text{raw}}-I_{\text{vCT}}\right),$$ (5)

where f is a 2D smoothing function. The scatter distribution is then subtracted from the raw intensity in order to create scatter-corrected projections as described in equations (3) and (6). Finally, the scatter-corrected projections are reconstructed with FDK into a scatter-corrected volume named CBCT_prior.

Figure 1.

Overview of the data processing performed by the prior-based scatter correction method of Park et al (2015), used as reference in this work.

2.3. Monte Carlo simulations

In order to train, validate and test the neural network, CBCT projections were simulated with Monte Carlo for all 48 patients using the GPU accelerated code MCGPU (Badal and Badano 2009). The geometry of an Elekta XVI system was implemented, using the same configuration as for the patients and phantom experimental images used in this study (centered panel, 20 cm collimator and no bowtie filter). The x-ray spectra was generated using the SpekCalc toolkit (Poludniowski et al 2009) for a tube voltage of 100 kVp, following the x-ray tube model used in Thing et al (2016). Projections were scored on a 1024 × 1024 grid with a 0.4 × 0.4 mm² pixel size, but downsampled to 512 × 512 with pixels of 0.8×0.8 mm² to limit memory usage.

MCGPU generates five outputs for each CBCT projection: the total intensity I_raw, the intensity associated with photons that did not experience scatter I_prim and the intensity from photons that experienced Compton, Rayleigh and multi-scatter (I_C, I_R and I_MS). The scatter distribution S associated with each projections can be calculated either by summing the last three components (S = I_C +I_R +I_MS), or doing the difference between the total and primary only images (S = I_raw −I_prim). A projection without any object in the field of view was simulated in order to generate the open field intensity I₀ to be used later in the process to normalize the scatter projection.

For each training and validation patient, 90 CBCT projections equally distributed over 360° were simulated, using 6×10⁹ photons per projection. Data augmentation was performed by flipping the projections horizontally and vertically, providing 360 projections per patient and a total of 13,680 projections to be used for training and validation combined. For the test patients, 540 projections equally distributed over 360° were simulated in order to reconstruct three image sets per patient: an uncorrected volume (CBCT_raw) reconstructed from raw projections $p_{\text{raw}}=-\ln \left(\frac{I_{\text{raw}}}{I_0}\right)$, a scatter-free volume (CBCT_SF) reconstructed from scatter free projections $p_{\text{SF}}=-\ln \left(\frac{I_{\text{prim}}}{I_0}\right)$ and a scatter-corrected volume (CBCT_NN) reconstructed from the scatter-corrected projections $p_{{\text{corr}}_{\text{NN}}}$, derived from the neural network as described in the following section. The proportions of simulated projections per patient in the training, testing and validation sets used in this study are similar to what was used by Maier et al (2019) and takes advantage of the fact that the training and validations sets do not have to be reconstructed to utilize data augmentation over simulation of additional projections, which get more and more similar for smaller angle increments. Finally, to evaluate the sensitivity of the trained model toward the accuracy of the x-ray spectrum model, an additional set of projections was simulated for one of the test patient, introducing an additional 2 mm aluminum filtration to the x-ray spectrum used during training.

2.4. Deep learning algorithm

The U-Net architecture used in this study follows the one from Maier et al (2019) and is presented in figure 2. The network is made of 7 layers separated into an encoding and a decoding parts, all connected by concatenations in the channel dimension. Each level of the encoding path applies three consecutive 3 × 3 convolutions followed by a parametric rectified linear unit (PReLU) (He et al 2015). At the end of each stage, the spatial resolution is reduced by two while the number of channel is doubled by applying a 3 × 3 convolution with 2 × 2 stride, also followed by a PReLU. The number of channels is thus augmented from 16 to 1024 along the encoding path. The decoding part is symmetrical to the encoding one, with the exception that the last convolution of each layer is replaced by a bilinear upsampling. As suggested by Maier et al, raw projection p_raw were chosen over raw intensity maps I_raw as input to train the network. However, contrary to Maier et al, we trained our network to predict the normalised scatter distribution $s=\frac{S}{I_0}$ instead of the pure scatter distribution S. This choice was made as projections created by the XVI system are gain corrected in order to assume a uniform open field intensity, while MC simulated scatter distributions includes the beam’s divergence.

Figure 2.

Deep convolutional neural network architecture used in this study.

The scatter corrected projections were obtained as follows. Combining equations (1), (3) and (4) one get:

$$p_{\text{corr}}=-\ln \left(e^{-p_{\text{raw}}}-\frac{\widehat{S}}{I_0}\right).$$ (6)

If ${\widehat{s}}_{NN}$ is the normalised scatter distribution predicted by the network from a given raw projection p_raw, the corrected projection $p_{{\text{corr}}_{\text{NN}}}$ is thus given by:

$$p_{{\text{corr}}_{\text{NN}}}=-\ln \left(e^{-p_{\text{raw}}}-{\widehat{s}}_{NN}\right).$$ (7)

To avoid artificially low intensity values, normalised scatter estimations made by the neural network were post-processed to yield a maximum of 95% of the normalised raw intensity, i.e. ${\widehat{s}}_{NN}=\min \left({\widehat{s}}_{NN},0.95\cdot e^{-p_{\text{raw}}}\right)$.

Projections were downsampled to 256× 256 before being sent to the network. Similarly, normalised scatter estimate had to be upsampled to the projection dimensions before performing scatter correction. The network was implemented in Pytorch (Paszke et al 2017) and the training was done on a NVIDIA TITAN Xp graphic card, using an Adam optimizer, a batch size of 4 and a learning rate of 5 × 10⁻⁶ for 150 epochs. The weights were initialized with a Glorot uniform initialization (Glorot and Bengio 2010), while the biases were initialized with zeros.

Two different loss functions were used to train the network: mean-square error (MSE) and mean absolute percentage error (MAPE). Those two quantities are calculated as follows:

$$\text{MSE}=\frac{1}{N}\sum _{\mathbf{d},n}{\left({\widehat{s}}_{NN}(\mathbf{d},n)-s(\mathbf{d},n)\right)}^2,$$ (8)

$$\text{MAPE}=\frac{100}{N}\sum _{\mathbf{d},n}\left|\frac{{\widehat{s}}_{\text{NN}}(\mathbf{d},n)-s(\mathbf{d},n)}{s(\mathbf{d},n)}\right|,$$ (9)

where d is the pixel coordinate, n is the sample index and N is the number of elements in each batch.

The impact of the output quantity that the network is trained to predict was also investigated. Following Hansen et al, we also trained a network to directly predict the scatter-free projection p_SF from p_raw. For this network, only MSE was used as cost-function. The networks trained to predict the normalised scatter distribution are referred to as p_raw → s while the network trained to predict scatter-free projections is referred to as p_raw → p_SF. In all scenarios, the accuracy of the scatter correction was evaluated by calculating the mean error (ME) and mean absolute error (MAE) of the reconstructed HU between CBCT_NN and CBCT_SF.

2.5. Phantom measurements

The applicability of the MC trained U-Net scatter correction to proton therapy dose calculation was evaluated using an anthropomorphic head phantom (Phantom Patient; The Phantom Laboratory, NewYork, NY) containing a real human skull. A total of 395 CBCT projections of the head phantom were acquired using the same imaging protocol as for the patients data. Here again, a 5× 5 median filter was applied to the projections before reconstructing the images with RTK, using the same Hamming window and truncation correction as for the patient data.

A CT volume of the head phantom was acquired to serve as reference in order to evaluate the accuracy of the scatter correction made by the MC trained neural network. External radio-opaque markers were used to reproduce the phantom’s alignment from the CT to the CBCT system and a final automatic rigid registration from the CT to the CBCT volume was additionally used to ensure optimal registration between the two image volumes. The automatic rigid registration was done with Plastimatch (Sharp et al 2010), using mutual information as similarity metric.

2.6. Dose calculation

Proton range accuracy achieved in the scatter corrected CBCT images was evaluated in the anthropomorphic head phantom using the CT volume as reference. A single posterior beam targeting a virtual brain tumor was optimized in the CT volume and re-calculated in the scatter corrected CBCT volume CBCT_NN as well as the uncorrected volume CBCT_raw. Dose calculation was performed in RayStation (v8.99, Raysearch, Sweden), using the IBA Dedicated Nozzle beam model and a 1.0 mm × 1.0 mm × 1.0 mm dose grid. Range accuracy was quantified by comparing the position of the distal 80% isodose line in the beam’s eye view on CBCT_NN, CBCT_raw and the reference CT volume. Only dose profiles achieving a value of at least 75% D_max were included in the range analysis, yielding a total of 1364 values of R₈₀ analysed for each modality.

For each test patient, an IMPT plan was created and optimized on the scatter free CBCT volume. The original contours from the planning CT were propagated to the simulated CBCT images to create the proton plans. A summary of the tumor locations and prescription doses used in this work is given in table 1. The IMPT plans were created in RayStation (v8.99, Raysearch, Sweden), using the IBA Dedicated Nozzle beam model, a 2.0 mm × 2.0 mm × 2.0 mm dose grid and a relative biological effectiveness (RBE) of 1.1, following our institution’s protocols. Multi-criteria optimization (MCO) was selected to optimize the plans, using the same objectives and constraints as the original photon plans, with the following exceptions: only one planning treatment volume (PTV) with uniform dose prescription was used per patient (i.e. high risk CTVs and boosts were not considered) and portions of the target situated below the shoulders were excluded, as the patient anatomy is truncated in those regions using a centered CBCT panel. An example of such a truncated CBCT and its impact on target volumes is provided in the supplementary material (stacks.iop.org/PMB/65/245022/mmedia).

Table 1.

Test	Tumor	Prescription
patient #	Location	Dose [cGy]
1	Mouth	6300
2	Sphenoid Sinus	5950
3	L. neck	6300
4	Tonsil	6000
5	Tongue	6300
6	Larynx	6000
7	Larynx	6000
8	Oropharynx	6000
9	Oropharynx	6000
10	Tongue	6300

3. Results

3.1. Network training

The loss curves associated with the training of the three networks considered in this study are presented in figure 3. Training took 21 h for both p_raw → s networks while it took 16.2 h for the p_raw → p_SF network. Iteration number 121 was used for the p_raw → s network trained using a MSE loss function, iteration 89 was used for the p_raw → s trained with the MAPE network and iteration 119 was used for the p_raw → p_SF network. Once the networks were trained, the average time to correct a single projection was 13.58 ms, which corresponds to a total of less than 5 seconds for a 360 projection CBCT volume.

Figure 3.

Values of the loss function for the training and validation patient sets using (a) the p_raw → s network with a mean-square error (MSE) loss function, (b) the p_raw → s network with a mean absolute percentage error (MAPE) loss function and (c) the p_raw → p_SF with a mean-square error (MSE) loss function.

3.2. Image analysis

Figure 4 presents the mean absolute error and mean error observed on the reconstructed CBCT_NN volumes of all test patients using the p_raw → s network trained with MSE and MAPE loss functions as well as the p_raw → p_SF network. Similar accuracy is observed for both p_raw → s networks, although a slightly superior performance is achieved using the MAPE loss function. The mean errors and mean absolute errors averaged over all test patients are (−0.801, 13.41) HU and (1.73, 15.48) HU for the p_raw → s network trained using the MAPE and MSE loss functions, respectively. A more important difference is observed between the p_raw → s and p_raw → p_SF networks, the latter achieving a noticeably lower HU accuracy. Mean absolute errors and mean errors averaged over all test patients for the p_raw → p_SF network are (−3.57, 20.23) HU respectively. Mean and mean absolute errors in the uncorrected images are −28.61 and 69.64 HU.

Figure 4.

Comparison of the (a) mean absolute error and (b) mean error obtained on CT numbers in CBCT images of all validation patients, corrected with the different networks considered in this study

In figure 5, reconstructed images from MC simulated CBCT projections are presented for a subset of the test patients cohort. Scatter free images (CBCT_SF) are compared to uncorrected (CBCT_raw) and scatter-corrected (CBCT_NN) images. Volumes were corrected using the p_raw → s network trained with the MAPE loss function. A noticeably superior agreement is found between CBCT_SF and CBCT_NN than CBCT_raw and the impact of the scatter correction is especially apparent in areas surrounded by bones.

Figure 5.

Representative CBCT images from a subset of the test patients cohort reconstructed from Monte Carlo simulated projections. The upper row shows the images reconstructed from scatter free projections, the second row shows the images reconstructed from raw projections and the third row shows the images reconstructed from scatter-corrected projections, using the trained U-Net. Differences between CBCT_raw and CBCT_SF and between CBCT_NN CBCT_SF are presented in the fourth and fifth row respectively. Slices are selected to show representative results of neck and base of skull regions.

A more quantitative evaluation of the accuracy of the scatter correction achieved with the neural network is provided in figure 6, where differences in HU between CBCT_SF, CBCT_raw and CBCT_NN are shown in a central slice of test patient # 6. A noticeably superior agreement is found between the scatter free image and the corrected one, compared to the uncorrected one. CT number deviations along a profile line are also presented, showing once again an almost perfect agreement between the scatter corrected image and the scatter free one.

Figure 6.

(a) Central slice of a Monte Carlo calculated scatter free CBCT (CBCT_SF), (b) difference between the corresponding uncorrected CBCT image (CBCT_raw) and CBCT_SF, (c) difference between the corrected CBCT image (CBCT_NN) and CBCT_SF (d) Profiles of CT numbers along the dashed line shown in (a).

3.3. Dose and range analysis

The impact of the U-Net based scatter correction on proton range accuracy is evaluated in figure 7. Proton range predictions made in uncorrected and scatter corrected real CBCT images of a head phantom (CBCT_raw and CBCT_NN) are compared to those made in a CT image of the same phantom. A substantially improved agreement is observed between the reference CT and CBCT_NN compared to CBCT_raw. In the central axis of the proton beam, a difference on R₈₀ between CBCT_NN and CT of 1.0 mm is observed. Over the whole beam, the root-mean square error on R₈₀ is 0.73 mm for CBCT_NN compared to 16.06 mm for CBCT_raw.

Figure 7.

(a) Reference dose distribution for the posterior proton beam calculated in the CT volume of the head phantom. Difference between the dose distributions calculated in the CT volume and the (b) uncorrected real CBCT image and the (c) scatter corrected CBCT volume. Depth dose curves along the dashed line shown in (a).

The 2%/2 mm gamma evaluation of CBCT_raw and CBCT_NN using CBCT_SF as reference is presented for the ten test patients in table 2. As expected, passing rates for CBCT_NN are substantially superior to those achieved with the uncorrected images. Dose calculations performed in CBCT_NN achieved passing rates between 94.18% and 100%, with a mean passing rate of 98.89%, compared to 68.44% for CBCT_raw.

Table 2.

2%/2 mm gamma pass rates for all test patients calculated in the uncorrected and scatter corrected CBCT volumes using the scatter free volume as reference.

Test patient #	1	2	3	4	5	6	7	8	9	10	mean
CBCTNN	99.92%	100%	100%	94.18%	100%	99.56%	98.21%	97.57%	99.47%	99.96%	98.89%
CBCTraw	69.18%	61.22%	65.94%	64.22%	70.32%	72.15%	66.55%	71.14%	73.12%	70.59%	68.44%

The robustness of the method toward the discrepancies between the x-ray spectra used for training and testing is evaluated for test patient #6 in figure 8. Dose differences between CBCT_SF and CBCT_NN are shown for two different cases, one where the spectra used for training and testing are the same and one where a 2 mm aluminium filtration is added to the spectrum used to generate testing images. The associated difference between CBCT_raw and CBCT_SF is also presented as reference. While one can see that the agreement between the doses calculated in scatter free and scatter corrected images is degraded when the x-ray spectrum is not perfectly modelled during training, the scatter correction is still a substantial improvement compared to the uncorrected images. The 2%/2 mm gamma evaluation passing rate for the altered x-ray spectrum situation is reduced to 98.56%, compared to 99.56% when the x-ray spectrum is perfectly modelled.

Figure 8.

(a) Reference dose distribution calculated in the scatter free CBCT volume for test patient #6. Difference between the dose distributions calculated in the scatter free CBCT and (b) the uncorrected CBCT, (c) the scatter-corrected CBCT (d) the scatter-corrected CBCT generated using a x-ray spectrum altered by a 2 mm Al filtration compared to the training data.

3.4. Comparison to an empirical projection based correction method

Figure 9 presents the dose distributions calculated in CBCT images of three representative patients corrected for scatter using the U-Net approach (CBCT_NN), using the prior-based method of Park et al (CBCT_prior) as reference. Small discrepancies can be observed between the two scatter correction methods, although they show generally good agreement. The gamma pass rates are (79.41%, 81.92%, 73.13%) and (98.24%, 99.12%, 98.79%) for the 2%/2 mm and 3%/3 mm criteria respectively.

Figure 9.

(a) Dose distributions calculated in CBCT images of three test patients. (b) Comparison between the dose distributions calculated in CBCT_NN and CBCT_prior.

4. Discussion

In this work, we evaluated the performance of a deep learning approach for projection-based scatter correction of CBCT data in the context of head an neck APT. A U-Net architecture was trained to reproduce Monte Carlo simulated scatter distributions in real time, following the approach of Maier et al (2019). A total of 48 head and neck patients were used to train, validate and test our network. The patient cohort size was sufficient to allowed us to test the network on ten patients which were not used for train or validation of the network. The applicability of the method to head and neck APT was evaluated by performing a comprehensive analysis on simulated patient data, as well as phantom measurements and real patient images.

In terms of image quality, we observed that the network architecture provided the best results when trained to predict the normalised scatter distribution $s=\frac{S}{I_0}$ instead of scatter free projections p_SF. Mean error and mean absolute error of (−0.80, 13.41) HU and (−3.57, 20.23) HU were obtained for the p_raw → s and p_raw → p_SF networks respectively. The impact of the loss function used to train the network, on the other hand, was shown to have a much lower impact on HU accuracy. The p_raw → s networks trained with MSE and MAPE loss functions yielded similar results, the latter yielding slightly superior performance.

The HU accuracy obtained in this study is similar to the one reported in Maier et al for head and neck patients, suggesting that our implementation of the network was satisfactory. However, it is worth mentioning that Maier et al only reported the mean error on CT numbers for two patients, in soft tissue regions only, making it hard to perform a more thorough comparison with our results. The mean error on HU achieved in this study (−0.8 HU) is also similar to the ones reported in similar studies making use of projection-based U-Net, such as in Hansen et al (−3.0 HU) and Landry et al (−2 HU). Much larger differences are however observed in terms of mean absolute differences, where values of 46 HU and 51 HU are reported for the test patients in Hansen et al and Landry et al respectively, compared to 13.41 HU in this work. This can be explained by the fact that both these studies used prostate images, where the magnitude of the scatter is expected to be larger and by the fact that they used networks trained to predict p_SF, which, according to our results, can provide higher MAE values in the reconstructed images.

Proton ranges calculated in CT and scatter corrected CBCT images of an anthropomorphic phantom showed good agreement, suggesting that a MC-based training of the network is suitable to correct real CBCT images for proton therapy dose calculation. The scatter correction allowed to reduce the RMS error on R₈₀ from 16.06 mm in uncorrected images to 0.73 mm. Considering a proton range of 11 cm, this represents an deviation of 0.66%, which is well within the range uncertainty typically considered during treatment planning. This range uncertainty is also consistent with the results reported in Thummerer et al, where a similar beam angle was used to evaluate range accuracy in CBCT images processed with an image-based U-Net trained to reproduce vCT images.

In terms of dose distributions accuracy, the IMPT plans used in this work had a 2%/2 mm pass rate above 99% for 7 of the 10 validation patients and a mean passing rate of 98.89%. This result is substantially superior to what was reported in Hansen et al where mean passing rates of 53% was observed using full IMPT. Using a single-field uniform dose (SFUD) optimization, Landry et al were able to achieve mean passing rates above 95%. Once again, the fact that both studies investigated prostate patients could explain why the scatter correction did not perform as well as in this work. In that regard, our results are once again in better agreement with those from Thummerer et al, where passing rates of 99.30% were achieved in head and neck patients. That study however only considered single-field plans, which are arguably more robust in terms of gamma analysis than the full IMPT multi-fields plans considered in our work.

This work is the first to evaluate the dosimetric impact of a U-Net trained to reproduce MC estimations of scatter distributions. While MC simulations can in theory estimate exact x-ray scatter distributions, their accuracy is influenced by the modelling of the CBCT system. Depending on the level of realism one aims to achieve, this task can be quite challenging. We evaluated the dosimetric impact of the accuracy of the spectrum modelling used to train and validate the network, by introducing a 2 mm aluminium filtration to simulate the projections of one test patient. Our results showed a slightly reduced dosimetric agreement with the scatter free scenario compared to the unaltered spectrum case. However, the 2%/2 mm gamma passing rate was only decreased by 1 percentage point, from 99.56% to 98.56%, which is still substantially superior to the 72.15% passing rate observed in the uncorrected volume. This suggest that the performance of the network is indeed influenced by the accuracy of the spectral modelling, although it is still somehow robust against moderate discrepancies such as the one we considered. It is worth mentioning that Maier et al demonstrated the feasibility of training only one network with different tube voltages to increase robustness toward spectral modelling, something we did not investigate in this study.

The applicability of our MC trained U-Net to real CBCT data was evaluated in patient images. Lacking a real ground truth for these cases, we used a thoroughly validated prior-based CBCT correction method as reference. This time, the 2%/2 mm gamma passing rate was lower than in the simulated data, with an average value of 78.15%. The difference in 2%/2 mm passing rates between the simulated and real CBCT images may be explained by an imperfect model of our CBCT system in our Monte Carlo simulations and by the fact that the reference prior-based method corrects for all low-frequency deviations in between I_vCT and I_raw. In addition to scatter detection, this includes some beam hardening artifacts such as cupping, which are not explicitly corrected in the simulation environment. Also, while this passing rate is noticeably lower than what was achieved using the simulated scatter-free images as reference, it is superior to what was reported for IMPT prostate treatments in Hansen et al despite the fact that their network was trained to reproduce the same prior-based CBCT correction used as reference. Even if CBCT scatter correction is conceivably more challenging in the context of prostate IMPT, this result suggest that a U-Net trained to reproduce MC simulated scatter distributions generalises well to real data. This is further supported looking at the more permissive 3%/3 mm criterion, which yielded a passing rate of 98.72% on average.

Globally, this study demonstrates the feasibility of CBCT scatter correction using deep convolutional neural networks in the context of head and neck APT. The trained network was able to estimate scatter distribution for a whole CBCT projection volume in a few seconds, achieving almost as good accuracy as full Monte Carlo. This processing speed is compatible with the needs of online APT and represents a substantial improvement over the reference prior-based method, which can require up to 10 minutes to correct a single CBCT volume (Park et al 2015). One advantage of the approach investigated in this work over similar U-Net based methods (Hansen et al 2018, Thummerer et al 2020, Landry et al 2019) is that it does not rely on potentially inaccurate DIR at any stage. This property is shared with unsupervised corrections methods such as GANs, which do not use paired data for training. However, one potential benefit of the proposed approach over GANs is that the complexity of the quantity learned by the ‘black box’ network is kept as low as possible. By learning the low frequency 2D scatter distributions and performing the subtraction in the projection domain, the current approach minimally alters the measured data and ensures a physically meaningful correction. Finally, since the network predicts a real quantity that can be estimated experimentally (Kyriakou and Kalender 2007, Wang et al 2010, Ouyang et al 2013), it could allow for easier commissioning and quality assurance than alternative deep learning based CBCT corrections.

Future work should investigate the generalisation of this approach to different anatomical sites in the context of APT. Maier et al demonstrated that a single network could be trained to perform scatter correction of head, thorax and abdomen scans, but this is yet to be confirmed for proton therapy dose calculation. Evaluation of the method for pelvis patients should also be performed, as other U-Net methods showed limited accuracy for IMPT dose calculation in this arguably more challenging anatomical site. It would also be interesting to evaluate how a similar method could be developed to correct for other types of artifacts, such as dental artifacts, which are usually more pronounced on CBCT images. Implementation of the method in the clinical workflow should also be investigated. This aspect might be more challenging for the proposed method than for image-based alternatives, as access to raw projections is not always granted on commercial systems. Finally, further studies should evaluate the impact of the scatter correction on additional APT tasks, such as auto-contouring or contour propagation.

5. Conclusion

In this work, we successfully implemented a U-Net architecture trained to perform projection-based scatter correction of CBCT data in the context of head an neck APT. The method achieved an accuracy similar to Monte Carlo simulations, while being considerably faster. An average 2%/2 mm gamma pass rate of 98.89% was obtained in full IMPT plans calculated on MC simulated CBCT images of 10 test patients. Range estimation in experimental CT and scatter-corrected CBCT images of an anthropomorphic head phantom yielded sub-millimetric agreement. The approach considered in this study also showed good agreement with a reference scatter correction method in patient CBCT patient images, while being one to two orders of magnitude faster.