Clinical use of iterative 4D-cone beam computed tomography reconstructions to investigate respiratory tumor motion in lung cancer patients

Abstract

Background. Cone beam computed tomography (CBCT) provides means for respiratory resolved volumetric imaging of the thorax. However, merely sorting the acquired projections into respiratory phases and performing a series of conventional three-dimensional (3D) reconstructions lead to clinically prohibitive reconstruction artifacts. This problem can be mitigated by iterative 4D reconstruction. We present a clinical evaluation of two iterative 4D-CBCT reconstruction algorithms during stereotactic body radiation therapy.

Material and methods. Two types of iterative 4D-CBCT reconstructions were performed utilizing: 1) total variation (TV) minimization; and 2) optical flow (OF) based deformable registration between phases. The reconstructions were initially evaluated on a lung phantom with a moveable target insert. Subsequently, 4D-CBCT reconstructions were performed for 19 patients on 2–3 CBCT projection datasets previously acquired for conventional 3D-CBCT reconstruction (∼650 half-fan projections per scan in a full one-minute gantry rotation). The 4D reconstructions were imported into a treatment planning system, where the gross tumor volume (GTV) was delineated and used to extract the tumor motion amplitude.

Results. For both phantom and patient scans, the iterative 4D-CBCT reconstructions had sufficient quality for GTV delineation when the breathing period was faster than 3.5 seconds (15 of 19 patients), but not for slower breathing periods (4 patients). The 3D tumor motion amplitude for the patients was significantly lower (p = 10⁻⁶, Wilcoxon signed rank test) in the OF reconstructions (mean 4.0 mm) than in the TV reconstructions (mean 5.3 mm).

Conclusion. TV and OF iterative 4D-CBCT reconstruction of the thorax in a lung phantom and for 19 patients was demonstrated from standard CBCT scans and used to estimate the daily lung tumor motion.

Tumors in the thoracic regions present special challenges due to breathing motion during the RT course [1]. In non-small cell lung cancer (NSCLC) patients, local failure rates are high and contribute to a low overall survival [2]. An increased radiation dose may improve the local control rate [3]. However, fatal toxicity has been shown to increase significantly in clinical dose escalation trials [3,4]. Therefore, tight margins to the target are essential and a measurement of the intra- and interfractional motion during the radiotherapy (RT) course is needed. For stereo tactic body radiotherapy (SBRT) this is of most importance for centrally located tumors [5,6] since high radiation doses are delivered to small tumors near critical organs in a few treatment fractions [7]. The expected range of respiratory motion, as obtained from a planning four-dimensional (4D) computed tomography (CT) scan (pCT), may be included in the planning target volume (PTV). However, breathing patterns may change between treatment planning and the treatment sessions [1]. Therefore, tumor positions and tumor motion should ideally be determined at each treatment fraction in order to adapt the treatment to variations in the breathing pattern.

3D-CBCT has the ability to visualize target positions at each treatment fraction and is therefore used for daily patient setup. An important application is the correction for baseline shifts immediately before treatment delivery [8,9]. 4D-CBCT provides respiratory phase resolved volumetric imaging of the thorax at the treatment session [10–12], allowing accurate determination of the respiratory motion. While 4D-CBCT is thus essential, simply extending conventional 3D-CBCT reconstruction algorithms to 4D by phase-by-phase reconstruction leads to severe reconstruction artifacts unless the total number of CBCT projections is substantially increased, e.g. by a slower gantry rotation or by making several gantry rotations [13–15]. The McKinnon-Bates (MKB) algorithm [16] reduces these artifacts by including a forward projection step. An initial image is reconstructed from the full data set, and a correction image is then added as a reconstruction from the difference between the measured raw data and the forward projected initial image using a conventional phase-correlated algorithm (FDK) [13]. A way to alleviate this problem without increasing the number of acquired x-ray projections is to use iterative reconstruction methods with appropriate regularization terms [17–20]. Despite publications of such iterative 4D-CBCT reconstruction algorithms in the past, they have not yet been studied clinically on a broader patient cohort. Thus, the aim of the present retrospective study is to translate two iterative 4D-CBCT reconstruction algorithms to clinical application. The first algorithm is based on total variation (TV) minimization and represents the most commonly proposed type of algorithm for iterative 4D-CBCT reconstruction [17]. Second, we evaluated a recently proposed reconstruction algorithm which utilizes optical flow (OF) based deformable image registration as a post-processing after an initial TV based reconstruction [20]. OF reconstruction facilitates reconstruction of each phase from the full set of acquired projections and therefore has a potential advantage over TV reconstruction in suppressing reconstruction artifacts caused by the small number of projections present for each respiratory phase. We investigated the possibility to delineate the gross tumor volume (GTV) in the 4D-CBCT reconstructions and used this to extract the lung tumor motion from the 4D-CBCT scans, which was compared with the tumor motion from the 4D-pCT scan.

Material and methods

4D-CBCT reconstructions

A GPU-based framework for iterative 4D-CBCT reconstruction from an undersampled set of CBCT projections was established providing the possibility of reconstructing conventionally acquired 3D-CBCT scans as 4D-CBCT [21]. TV and OF iterative 4D-CBCT reconstructions were performed with an isotropic voxel size of 1.2 mm and 10 respiratory phases. The mathematical details of the tested reconstruction algorithms can be found in reference [20]. For the scope of this paper we provide a brief non-mathematical description of the key concepts: Both algorithms are based on minimizing a dedicated cost function. Conceptually we traverse the infinite set of all possible 4D reconstruction, synthesizing x-ray projections for each reconstruction, to quantify which reconstruction most likely produced the measured data. Due to the temporal binning (undersampling) of the projection data many reconstruction, with very different characteristics, conform to each phase's set of measured projections. An additional regularization term (quality measure) is thus introduced to specify which reconstruction to choose from the conforming set. Different regularization terms are used in the TV and OF reconstructions. TV promotes piecewise constant reconstructions and hence suppresses noise and unstructured aliasing in the reconstructions. The OF algorithm regularization term is instead based on image registration, allowing each temporal phase to be resampled into all other phases. Thus for each temporal phase, despite the undersampling caused by temporal binning, this enables x-ray synthesis of all acquired projections. Consequently, the OF regularization term promotes consistency versus all acquired projections for each temporal phase. The reconstructions were imported into the treatment planning system (TPS) (Eclipse, Varian Medical Systems) in DICOM format for subsequent analysis.

Phantom CBCT scans

Experimental validation of the 4D-CBCT reconstruction methods was performed using a 008A Dynamic Thorax Phantom [Computerized Imaging Reference Systems, Inc. (CIRS), Norfolk, VA] with a moveable lung equivalent rod with an embedded spherical solid water target insert (30 mm diameter). Half-fan 3D-CBCT scans of the phantom were acquired during a one-minute full gantry rotation (on average 650 projection images) with an On-Board Imager system (OBI) (Varian Medical Systems). CBCT scans were acquired with a cos⁴ target motion in the superior-inferior (SI) direction with 6 and 12 mm peak-to-peak amplitude, and periods of 2.5, 3, 4 and 6 seconds to investigate the impact of motion amplitude and period on the 4D-CBCT reconstructions. The CBCT scans were started at random times relative to the phantom motion. The phantom CBCT projections were sorted into 10 phases by first finding the projections with maximum inspiration and subsequently subdividing each breathing cycle into 10 phase bins of equal temporal length. The maximum inspiration projections were identified by scrolling through the projections to point out the images with the most inferior target position. The target was not visible in one third of the projection images due to the laterally shifted imager position of the half-fan CBCT mode. Here, knowledge of the periodic target motion was used to determine the maximum inspiration projections. Both the TV and OF 4D-CBCT reconstructions were imported into the TPS, where the GTV was delineated in the mid-ventilation (mv) phase. 4D-pCT scans of the phantom with a SI target motion of 6 and 12 mm peak-to-peak amplitude, and periods of 2.5 and 3 seconds were performed for comparison with the 4D-CBCT reconstructions with a known ground truth of the target motion. The TV and OF reconstructions of the phantom were visually compared with MBK, a 4D FDK algorithm.

Patients and treatment planning

We retrospectively selected 19 consecutive lung SBRT patients (median age of 75 years, range 60–82 years) treated at Aarhus University Hospital between April and August 2012. Supplementary Table I (available online at http://informahealthcare.com/doi/abs/10.3109/0284186X.2014.927585) summarizes key characteristics of the patient group. Radiotherapy doses of 67.5 Gy were delivered in three fractions. The GTV volume of the tumors had a median volume of 8.4 cm³ (range 2.9–55.4 cm³). For two patients with two tumors (two primary tumors and two colon cancer metastasis, respectively), only one of the tumors were selected for 4D-CBCT motion analysis.

A free-breathing 4D-pCT was acquired using an optical breathing signal to sort the acquired images into 10 respiratory phases. The mv phase was manually selected and the GTV delineated in the TPS. The median time interval between pCT scan and treatment start was 9 days (range 4–18 days). At each treatment session, a 3D-CBCT scan for patient positioning was acquired during a one-minute full 360° gantry rotation (on average 650 projection images) using the same half-fan CBCT mode as for the phantom scans. The CBCT projections were saved and afterwards used for offline 4D reconstruction. For three patients, CBCT projections were only available for two of the three fractions. A total of 54 CBCT scans were thus investigated in this study.

Phase sorting

Similar to the phantom scans, the CBCT projections were sorted into 10 respiratory phases by identifying the projections corresponding to maximum inspiration. For patients whose diaphragm was visible in all projections, the phase sorting was done using the Amsterdam Shroud method [22]. For scans in which the diaphragm was not visible in all projections, the maximum inspiration projections were instead identified by analyzing the fluctuation of the image intensity in the lung projection images [23]. For a few patients, this method of automatic phase sorting also failed and the maximum inspiration projections were identified manually by two independent physicists by scrolling through the projections. For comparison, the CBCT scans sorted with the Amsterdam Shroud were also sorted by the lung density fluctuation method.

Internal target motion

The GTV was delineated in the mv phase for the 4D-CBCT reconstruction, which was selected identical to the mv phase selected in the 4D-pCT scan. Deformable registration in the Smart Adapt software package (Varian Medical Systems) was performed between the mv phase and the remaining nine phases of both the 4D-pCT and the 4D-CBCT with a demons algorithm. This registration was used to propagate GTV to all phases of the 4D-pCT and -CBCT reconstructions. The signature of the tumor and the GTV structure was verified visually. The center-of-mass positions of GTV in the 10 phases were used to extract the tumor motion amplitude in each pCT and 4D-CBCT scan.

Results

Phantom CBCT scans

As seen in Figure 1, the reconstruction quality was superior for the phantom CBCT scans acquired with shorter target motion periods. Significant artifacts are present in the reconstruction corresponding to the longest respiratory cycles, making GTV delineation and deformable propagation possible. Similar results were obtained with the target motion of 6 mm (not shown). Figure 2 compares the target motion amplitude determined from the 4D-CBCT reconstructions with a target motion period of 2.5 and 3 seconds with the corresponding motion amplitudes extracted from the 4D-pCT scans. Also shown in Figure 1a are the MBK reconstructions of the phantom, which show a high image quality of the stationary parts of the phantom. Figure 1b shows a corresponding Figure in sagittal view, i.e. the direction of the motion, where it is seen how the MKB algorithm fails to correct for the motion artifacts, whereas the iterative TV and OF reconstruction do correct for the motion artifacts.

Figure 1. MKB, TV and OF 4D-CBCT reconstructions of the phantom from CBCT scans with target motion of 12 mm and a period of 2.5, 3, 4 and 6 seconds of the mv phase in (a) transversal view and (b) sagittal view.

Figure 2. 3D amplitude of 4D pCT versus 4D-CBCT with the TV and the OF reconstruction for phantom CBCT scans with target motion amplitudes of 6 and 12 mm and periods of 2.5 and 3 seconds.

Phase sorting

Amsterdam Shroud phase sorting was possible and therefore used for 24 of the 54 patient CBCT scans. Density fluctuation sorting was possible for the same scans plus 22 additional scans with only partly visible diaphragm. Manual sorting was performed for the remaining eight CBCT scans. For CBCT scans where both the Amsterdam Shroud method and the density fluctuation method were performed, the projections corresponding to maximum inspiration varied at most one projection between the two methods. For the eight scans where phase sorting was performed manually, the projections defined to correspond to maximum inspiration varied 0–2 frames between the two independent physicists.

Image quality of clinical iterative 4D-CBCT reconstructions

The image quality of the iterative 4D-CBCT from both TV- and OF-based 4D-CBCT reconstruction was suitable for GTV delineation and thereby tumor motion estimation in 15 of 19 patients. To illustrate the obtained image quality, Figure 3 depicts the 4D-CBCT reconstructions for a patient in which GTV delineation was possible. For four patients who had breathing periods longer than 3.5 seconds, the artifacts in the 4D-CBCT reconstructions were too severe for reliable tumor motion estimation. Figure 3 shows an example of such an unsatisfactory 4D-CBCT reconstruction of the mv phase.

Figure 3. Examples of TV and OF 4D-CBCT reconstructions for patient 5 and patient 12. A transversal, coronal and sagittal view of the mv phase is shown. The tumor is visible for patient 5, while for patient 12 the tumor is impossible to visualize due to reconstruction artifacts.

Internal motion

The volume of the deformable propagated GTV in the 10 respiratory phases varied in the same range when comparing the 4D-CBCT reconstructions with 4D-pCT. The quantified tumor motion was largest in the SI direction: 4D-pCT (range 0.5–9.9 mm), 4D TV CBCT (range 0.7–11.0 mm), 4D OF CBCT (range 0.5–9.0 mm).

The mean 3D peak-to-peak amplitudes were 5.4 mm (range 2.4–12.0 mm), 5.3 mm (range 1.7–12.3 mm) and 4.0 mm (range 1.6–9.9 mm) for the 4D-pCT, 4D TV CBCT, and 4D OF CBCT, respectively. The 3D tumor motion amplitude was significantly lower (p = 10⁻⁶, Wilcoxon signed rank test) in the OF reconstructions (mean 4.0 mm) than in the TV reconstructions (mean 5.3 mm) (Figure 4a). There was a large standard deviation of the 3D motion amplitude extracted from the 4D TV CBCT (Figure 4b), while the 4D OF CBCT shows slightly smaller 3D motion amplitude than observed in the pCT (Figure 4c).

Figure 4. 3D amplitude for 2–3 fractions for the 15 patients in which, 4D-CBCT reconstruction was possible. (a) 4D TV CBCT versus 4D OF CBCT, (b) 4D-pCT versus 4D TV CBCT, and c) 4D-pCT versus 4D OF CBCT. In b and c points related to different fractions are connected by vertical lines.

Discussion

Iterative 4D-CBCT reconstructions were performed for both a physical phantom and for 54 conventionally recorded scans for 19 SBRT patients. Automatic respiratory phase sorting based on either the Amsterdam Shroud or the density fluctuation method was in agreement within one projection image, consistent with prior results from Shieh et al. [24]. 4D-CBCT were reconstructed using both the TV and OF methods. For 15 of 19 patients it was possible to delineate the GTV and extract the tumor motion from the 4D reconstructions providing phase resolved volumetric imaging of the thorax in relation to each treatment fraction. A study by Rit et al. [25] compared tumor motions between 4D-pCT and 4D-CBCT with and without motion correction. They assumed that the breathing motion patterns did not vary in time, by making a registration on the 4D-pCT and propagate this to the 4D-CBCT reconstructions, which is different from our point of view. In our method each CBCT is self-contained, and we do not assume that the tumor motion to be stable in time during the treatment course. Rit et al. found tumor motion amplitudes of similar magnitude as our tumor motion amplitudes. They also found that the registration errors increased with tumor motion amplitude.

Although the iterative CBCT reconstructions were successful for most patients, there is still a lack of robustness in the algorithms. As seen in Figure 3 the 4D-CBCTs contained clinically prohibitively artifacts in four patients exhibiting slow breathing. These observations from the clinically data were confirmed in the phantom measurements, as shown in Figure 1. Generalizing our results obtained from two selected reconstruction algorithms for iterative 4D-CBCT, reconstruction from a one-minute single-rotation scan is thus only useful if the breathing period is shorter than 3.5 seconds. A comparison of the TV and OF reconstructions with the MKB reconstructions showed that the MKB yields a small number of artifacts, since the values in the difference in the correction reconstruction are usually smaller in magnitude than in the original raw data. As seen in Figure 1a, the MKB reconstructs images of a high quality in stationary areas, whereas it is challenging for this algorithm to correct for those motion artifacts, such as streaks and blurring, that are already included in the initial image, as seen in Figure 1b. The iterative reconstructions algorithms were superior in correction of the motion artifacts.

The projection images in each phase bin are distributed in bundles of consecutive projections that are separated by gaps, where the other nine bins are recorded. For the applied half-fan CBCT mode and a 3.5 seconds breathing period, the projections that include the tumor will be distributed in 8–9 bundles, each spanning 2.1° and being separated by 21°. The longer the breathing period, the longer is the angular separation between these projection bundles. This causes the distribution of projection angles to be more non-uniform and consequently the reconstruction problem to be more ill-posed numerically. In turn, a reduction of the obtained image quality is observed. There are two obvious ways to alleviate this problem and hereby ensure that the distribution of projections angles becomes more uniform for patients with slow breathing. The first option is to rotate the gantry with reduced speed during the CBCT acquisition. Alternatively, it is possible to acquire data over multiple gantry rotations at a reduced acquisition rate. Both strategies could be set up to acquire an identical amount of projections as for the single one-minute rotation protocol that was applied for the data in this study, i.e. they would not result in increased radiation dose to the patient. Instead, the scan time would be increased. Rit et al. [25] confirm empirically that CBCT acquisition over 4 minutes results in better image quality than a one-minute acquisition due to a spread out of the projections.

For 15 patients it was possible to delineate the GTV and extract the tumor motion from the 4D-CBCT reconstructions. The tumor motion found in the 4D-CBCT reconstructions from TV showed large variations while the tumor motion from the OF reconstructions was substantially more stable throughout the treatment course (Figure 4), and it was found that the 3D tumor motion amplitude was significantly lower the OF reconstructions than in the TV reconstructions.

An initial deformable registration between the phases from the TV reconstruction is required for an OF reconstruction and determines its outcome. An accurate registration should theoretically result in a OF reconstruction which improves the quality of the underlying TV reconstruction as the OF reconstruction is computed from the full set of projections. OF reconstruction can thus be considered as a post-processing step to the TV reconstruction. A poor initial registration of the TV phases on the other hand would cause unpredictable results for the subsequent OF reconstruction. It is important to emphasize that the two algorithms are not independent reconstruction alternatives, a good TV reconstruction and registration hereof is a prerequisite for a satisfactory OF reconstruction subsequently. The most likely explanation for the smaller motion amplitudes we observed in the OF reconstructions is an underestimation of the tumor motion in the underlying TV phase registration. For OF registration this could be an effect of over-regularization and consequent damping of the tumor motion. The TV reconstruction does not have this limitation but other drawbacks instead. It promotes piecewise constant reconstructions whereby: 1) small structures, such as the vessels in the lung could disappear from the reconstruction entirely; or 2) small cavities with lower/higher intensities as their surrounding get incorrectly filled with an intensity matching of its surroundings. In general, we observed that the OF reconstructions improved the qualitative impression of the reconstructions, in particular in the stationary parts of the tissue. As we demonstrated, however, motion tends to be underestimated for the OF reconstructions. Improved results from the initial deformable image registration are required to improve this aspect.

In conclusion, this translational study demonstrated iterative TV minimizing and OF registration-based 4D-CBCT reconstruction of the thorax at a lung phantom and 2–3 treatment fractions for 19 patients from a standard CBCT scan. The image quality of the 4D-CBCT reconstructions were sufficient in 15 patients for GTV delineation and made it possible to quantify the tumor motion of most patients at all treatment fractions.

Declaration of interest: The authors report no conflicts of interest. The authors alone are responsible for the content and writing of the paper.

This work was supported The Danish Cancer Society and CIRRO – The Lundbeck Foundation Center for Interventional Research in Radiation Oncology.