Progressive cone beam CT dose control in image-guided radiation therapy

Abstract

Purpose: Cone beam CT (CBCT) in image-guided radiotherapy (IGRT) offers a tremendous advantage for treatment guidance. The associated imaging dose is a clinical concern. One unique feature of CBCT-based IGRT is that the same patient is repeatedly scanned during a treatment course, and the contents of CBCT images at different fractions are similar. The authors propose a progressive dose control (PDC) scheme to utilize this temporal correlation for imaging dose reduction.

Methods: A dynamic CBCT scan protocol, as opposed to the static one in the current clinical practice, is proposed to gradually reduce the imaging dose in each treatment fraction. The CBCT image from each fraction is processed by a prior-image based nonlocal means (PINLM) module to enhance its quality. The increasing amount of prior information from previous CBCT images prevents degradation of image quality due to the reduced imaging dose. Two proof-of-principle experiments have been conducted using measured phantom data and Monte Carlo simulated patient data with deformation.

Results: In the measured phantom case, utilizing a prior image acquired at 0.4 mAs, PINLM is able to improve the image quality of a CBCT acquired at 0.2 mAs by reducing the noise level from 34.95 to 12.45 HU. In the synthetic patient case, acceptable image quality is maintained at four consecutive fractions with gradually decreasing exposure levels of 0.4, 0.1, 0.07, and 0.05 mAs. When compared with the standard low-dose protocol of 0.4 mAs for each fraction, an overall imaging dose reduction of more than 60% is achieved.

Conclusions: PINLM-PDC is able to reduce CBCT imaging dose in IGRT utilizing the temporal correlations among the sequence of CBCT images while maintaining the quality.

Cone beam CT (CBCT) has been routinely used in image guided radiation therapy (IGRT) to acquire a latest volumetric image of the patient's anatomy for precise treatment setup.^{1, 2, 3, 4, 5} However, compared with the treatment dose that is localized to the tumor area, considerable imaging dose is delivered to healthy organs from repeated CBCT scans,^{3, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17} elevating biological risks such as secondary cancer and genetic defects. Hence, the daily use of CBCT in IGRT has been limited, despite its apparent advantages for image guidance. Developing novel CBCT technologies with reduced imaging dose is therefore critically necessary. Over the recent years, a great amount of research effort has been devoted to this topic, such as iterative reconstruction approaches based on physics models^{18, 19, 20, 21, 22, 23, 24, 25} or the compressed-sensing theory.^{26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43}

Compared to many CBCT applications in other clinical scenarios, one unique feature of CBCTs in IGRT is that the same patient is repeatedly imaged during a treatment course. Although many existing dose-reduction techniques can be applied to CBCT in IGRT, they treat each scan independently, and the same technique is used for all fractions. Yet, the interfraction variation of CBCT image contents is expected to be small for a patient. By exploiting this unique correlation along the temporal dimension, one can, in principle offer even further dose reduction for CBCT in IGRT, in addition to those existing low-dose techniques.

In light of this fact, we propose in this Letter a progressive dose control (PDC) scheme for CBCT in IGRT. Specifically, as opposed to the current CBCT practice using a static scan protocol, a dynamically adjusted protocol is applied, which gradually reduces the imaging dose at each treatment fraction. The CBCT quality is maintained by incorporating previously available images as prior knowledge via a prior-image based, nonlocal means (PINLM) method. Over the treatment course, the increasing amount of prior knowledge from previous scans helps prevent a drop in image quality due to dose reduction. This allows for progressive dose reduction at each fraction, yielding a significant overall reduction in imaging dose over the entire treatment course.

Let us first present the PINLM method. We denote the CBCT image acquired at the day k + 1 in a treatment course under a low-dose protocol as $f_{k+1}^{\mathrm{Ori}}$. PINLM will process this image by incorporating the redundant information existing in all previous prior CBCT images, denoted as f_l(l = 1, …, k), as follows:

$$\begin{array}{c}\hfill f_{k+1}\left(\mathit{x}\right)=\frac{1}{Z}\left[{\sum }_{l=1}^k{\sum }_{\mathit{y}}\left(w_{l,\mathit{y}}·f_l\left(\mathit{y}\right)\right)+w^{\mathrm{Ori}}·f_{\mathrm{k}+1}^{\mathrm{Ori}}\left(\mathit{x}\right)\right],\end{array}$$ (1)

where f_{k + 1}(x) represents the enhanced image voxel value at x; w^Ori and w_{l, y} are weighting factors of f^Ori(x) and f_l(y). Specifically, w^Ori is a constant and w_{l, y} is estimated by the similarity between the voxel x in the current image $f_{\mathrm{k}+1}^{\mathrm{Ori}}$ and the voxel y in the prior image f_l given by $w_{l,\mathit{y}}=\exp \left\{-\frac{1}{h^2}·{\Vert V^{\mathrm{Ori}}\left(\mathit{x}\right)-V_l\left(\mathit{y}\right)\Vert }_2^2\right\}$. w^Ori = 0.5 is used throughout this paper. Decent image quality can be obtained with this empirically chosen value in all cases studied in this work. Here, V^Ori(x) and V_l(y) are small cubic volumes centering at x in $f_{\mathrm{k}+1}^{\mathrm{Ori}}$ and at y in f_l. By comparing the similarity between these two cubes in this way, large w_{l, y} values are automatically assigned to those voxels y that are similar to x. In this process, h is a tuning parameter reflecting the rigor level of the similarity criteria, and $Z={\sum }_{l=1}^k{\sum }_{\mathit{y}}{w}_{l,\mathit{y}}+w^{\mathrm{Ori}}$ is a normalization factor. The proposed PINLM algorithm stems from the original NLM method for image processing.⁴⁴ In contrast to original NLM borrowing information from the same image, PINLM fetches voxel values from both prior and current images. Note that if this processing scheme is utilized in every fraction, all images f_l for l = 1, …, k will have already been updated prior to the fraction k + 1, and will be of relatively high quality. Their utilization is beneficial for processing the image at fraction k + 1. While image enhancement using NLM has also been used in diagnostic CT imaging,^{45, 46} our scheme distinguishes itself by its dedicated design for IGRT, where a number of prior CBCT images are readily available.

Along with this PINLM method, we propose a PDC scheme. Specifically, imaging dose at each fraction is gradually reduced within the treatment course. The resulting CBCT image is processed by the PINLM method using all previously available CBCT images to enhance its quality. Note that this method does not specify a particular reconstruction method and it is thereby compatible with conventional FDK-based reconstruction method, as well as advanced iterative reconstruction techniques. The workflow is shown in Fig. 1a.

Figure 1.

(a) The workflow of PINLM-PDC. Illustrations of the relationship between imaging dose and image quality in (b) using the proposed PINLM-PDC scheme and in (c) using the current standard low-dose scheme. Plots are drawn for illustrative purpose and not to scale.

Given the expression in Eq. 1, we can explicitly compute the variance of the voxel x at f_k+1. If we require that the voxel variance be maintained at a constant level of σ² for each fraction k, the allowable variance ${\sigma }_{k+1}^2$ in the original image $f_{\mathrm{k}+1}^{\mathrm{Ori}}$ before PINLM processing can be derived as ${\sigma }_{k+1}^2={\sigma }^2\left[{\left(k{\sum }_{\mathit{y}}{w}_{\mathit{y}}+w^{\mathrm{Ori}}\right)}^2-k{\sum }_{\mathit{y}}{w}_{\mathit{y}}^2\right]/{w^{\mathrm{Ori}}}^2$, where $w_{\mathit{y}}$ is the weighting factor of f_l(y) for l = 1, …, k Here, we neglect the dependence of $w_{\mathit{y}}$ on the image but treat them as known constants. We also neglect the possible correlations among CBCT images at different fractions due to the PINLM processing. Considering that the imaging dose on the fraction k + 1 is approximately proportional to mAs, which is inversely proportional to ${\sigma }_{k+1}^2$, we arrive at an expression for the required dose at the k + 1 fraction $D_{k+1}\propto {w^{\mathrm{Ori}}}^2/\left[{\left(k{\sum }_{\mathit{y}}{w}_{\mathit{y}}+w^{\mathrm{Ori}}\right)}^2-k{\sum }_{\mathit{y}}{w}_{\mathit{y}}^2\right]$, in order to maintain the image quality in terms of noise level. Using a set of parameters of w^Ori = 0.5 and $w_{\mathit{y}}$'s from a sequence of realistic patient CBCT images, we can plot a dose reduction curve that is monotonically decaying, as illustrated by the imaging dose (theoretical) curve in Fig. 1b. In contrast to the current practice of using a constant dose level to maintain the image quality [Fig. 1c], this PINLM-PDC allows for a gradual dose reduction along the course of treatment.

We would also like to comment on the validity of this simple analysis. This analysis assumes that, for a voxel x, corresponding similar voxels y in other prior images can be successfully found. In practice, this may not hold due to the excessive noise when the dose is reduced to a certain extent. Starting from the point where similar voxel identification is severely impeded, the efficacy of PINLM is compromised, leading to a saturation of the dose reduction curve, as illustrated by the dashed curve in Fig. 1b.

One practical issue of the proposed PINLM-PDC scheme is its high computational burden. To speed up the computation, we implement the PINLM module in CUDA according to Ref. 47 and run it on an NVIDIA Tesla C1060 Graphics Processing Unit (GPU) to utilize its enormous parallel-processing capability. For each voxel, we have also limited the search range for similar voxels in previous images to a cubic volume of 11 voxels in each side. The cube size around each voxel is 3 × 3 × 3 voxels. Other details regarding this implementation are similar to what is described in Ref. 47. All CBCT images before PINLM are reconstructed by the FDK algorithm.⁴⁸

Our first evaluation is conducted using real experimental data of a Catphan phantom scanned with a Varian onboard imaging CBCT system (Varian Medical System, Palo Alto, CA). The phantom is deliberately placed at slightly different locations for each scan in an attempt to mimic day-to-day variations. The magnitude of these shifts is ∼7 mm along the superior-inferior direction and ∼2 mm along each lateral direction. The scans are carried out with successively decreased mAs levels of 0.8, 0.4 [the current standard low-dose protocol, recommended by AAPM TG75 (Ref. ⁹)] and 0.2 mAs. As shown in Fig. 2, compared to the result under 0.4 mAs with a noise level of 23.03 HU, the original image at 0.2 mAs has an amplified noise level of 34.95 HU. Yet, the PINLM method improves the quality considerably by incorporating prior information from the scan at 0.4 mAs, yielding a noise level of 12.45 HU, which is than that in a regular image in a regular image scanned at a high dose level of 0.8 mAs. To investigate whether or not there is any resolution degradation, we profile the intensity values across the finest distinguishable line pairs in the resolution slice of the phantom [Fig. 3a]. We also plot in Fig. 3b the profiles of the ramp wire in the contrast slice, indicated by the arrow in Fig. 2 and measure effective slice thickness based on the full width at half maximum (FWHM). The profile comparisons clearly show that the image resolution is well maintained. The measured effective slice thicknesses agree within 0.05 mm among all cases.

Figure 2.

CBCT images scanned at (a) 0.8 mAs, (b) 0.4 mAs, (c) 0.2 mAs, and (d) enhanced by using PINLM on image (c) with image (b) as the prior image. The noise standard deviation (std) is measured inside the square area and the averaged std is labeled in the upright corner of each subfigure. (Top row) Contrast slice displayed with window [−125, 225] HU and (bottom row) resolution slice displayed with window [−225, 725] HU.

Figure 3.

Image profiles at resolution line pairs (a) and the ramp wire (b) for different cases. For a better visual comparison, profiles are shifted vertically by 200 HU each.

The proposed PINLM-PDC scheme is further evaluated with synthetic patient data. A head-and-neck (HN) cancer patient's CT image is used as input. The CT is deformed in the soft-tissue area, especially around the esophagus and bronchia cavity, by using manually generated smooth vector fields. The vector fields are different in each day, to mimic the intrafractional anatomical variations. X-ray projections with different noise levels are generated by using a hybrid Monte Carlo simulation toolkit gDRR.⁴⁹ Two sets of CBCT scans are simulated for four consecutive fractions. In the first set, the current standard low-dose protocol (0.4 mAs) (referred as standard protocol hereafter) is applied to all fractions. In the second set, gradually reduced dose levels of 0.4, 0.1, 0.07, and 0.05 mAs are used in the four fractions (referred as dynamic-dose protocol), respectively. As shown in Fig. 4, while the standard protocol maintains a constant image quality, the CBCT images are degraded by increasing noise in the dynamic-dose protocol. In particular, low-contrast objects that are important for soft-tissue delineation in IGRT (Ref. ⁵⁰) are more vulnerable to the increased noise, as indicated by the arrows. However, under the PINLM-PDC framework (referred as PINLM-PDC), the image quality is maintained at a satisfactory extent. Note that the image of day 4 still shows inferior low-contrast preservation to a certain extent, when compared to the image using the standard protocol, indicating that 0.05 mAs may be the limit for dose reduction in this PINLM-PDC scheme, i.e., at this level the dose reduction curve in Fig. 1b saturates. Considering that the imaging dose is approximately proportional to mAs, we find an overall reduction in dose of more than 60% in the first four days when using the PINLM-PDC scheme.

Figure 4.

Synthetic CBCT images of the same longitudinal position on days 1–4. (a) Standard protocol; (b) dynamic-dose protocol (gradually reduced exposure level at 0.4, 0.1, 0.07, and 0.05 mAs for days 1–4); (c) PINLM-PDC with the same mAs settings as in (b).

To investigate the results in detail, we display the difference between the standard-protocol (0.4 mAs) and the dynamic-dose protocol images [Fig. 5a], as well as the difference between the standard-protocol and the PINLM-PDC images [Fig. 5b]. In the dynamic-dose protocol case, a large amount of noise is seen throughout the images. However, the noise level is greatly reduced by the PINLM algorithm. Interestingly, the noise signal shows spatial structures. To demonstrate that the structure is not due to, for example, incorrect edges brought by the algorithm from prior images, we also compare some intensity profiles on the three image sets in Figs. 5c, 5d, 5e. It is clearly observed that the edges are well aligned among those curves.

Figure 5.

(a) The difference (in days 2–4) between the standard-protocol (0.4 mAs) and the dynamic-dose protocol images. (b) Difference between the standard-protocol and the PINLM-PDC images. (c)–(e) show image intensity profiles along the cuts indicated by the arrows.

In this Letter, we propose a practical CBCT dose control scheme (PINLM-PDC) for IGRT. Our proof-of-principle experiments demonstrate its feasibility. PINLM-PDC does not depend on any specific reconstruction algorithm, and, hence, is highly compatible with current clinical systems. It does not require access to raw projection data or any hardware modification. This technique is expected to dramatically reduce the dose induced by frequent CBCT scans.

We would also like to briefly discuss a few issues related to the clinical implementation of our algorithm. First, the success of this algorithm relies on the ability of a voxel to find its similar voxels in prior images and use that intensity information to enhance itself. A rigid registration is hence necessary to align the patient anatomy with that in the prior images, to facilitate the searching procedure. This is particularly important when a small-size search window is used for efficiency considerations. In addition, future studies need to be conducted to determine the best search window size, which could even be chosen adaptively. Second, there is always a concern regarding the validity of using prior images, for instance, about the risk of introducing incorrect edge information to the new image. We have demonstrated the capability of PINLM-PDC in this respect through the simulated patient studies; however, further investigations on real patient data are required in this direction. Another important question to answer is how many prior images should be used in this algorithm, as more images may considerably slow down the computation with diminishing gains in image quality. Third, the image dose allocation of 0.4, 0.1, 0.07, and 0.05 mAs was just used for proof-of-concept. The optimal dose allocation scheme under the PINLM-PDC framework is an important topic. Finally, parameter optimization in PINLM is also a critical step. All of these issues will be investigated in our future studies.