Automated target misalignment correction for cone beam computed tomography-based online adaptive radiotherapy of locally advanced lung cancer patients.

1
Introduction
Patient anatomy can change over the course of radiotherapy, leading to a sub-optimal dose delivery due to differences with respect to the anatomy represented in the original treatment plan. In the case of lung cancer patients, posture changes, atelectasis, infiltrations, tumor regression, or rigid tumor baseline shifts could negatively impact the treatment quality [1]. In particular, differential motion between pulmonary primary tumor (GTVprim) and mediastinal lymph nodes is frequently observed in locally advanced lung cancer patients (LALCP), leading to enlarged treatment margins [2].
In general, LALCP are treated using a daily image guided radiotherapy (IGRT) protocol with cone beam CT (CBCT). In-room daily patient alignment can be performed with a CBCT-to-planning CT (pCT) carina registration, used as surrogate for the mediastinal lymph nodes position according to a previous study on treatment table corrections [3]. Due to differential motion between GTVprim and lymph nodes, residual GTVprim misalignments remain after patient alignment.
Adaptive radiotherapy monitors anatomical changes over the course of treatment and restores dose conformity by modifying the treatment plan according to the new anatomy [4,5]. This process requires significant resources, given that it requires modifications of delineations, as well as a 3D image that represents the new anatomy and that is suitable for subsequent dose calculation. Online adaptive radiotherapy (OART) aims at adapting the plan immediately before the treatment delivery [6]. Workflows involving OART demand fast preparation of an adapted plan within a few minutes. To this end, automated tools that enable OART need to be established [5,7,8].
In-room CBCTs, commonly used for patient positioning, give insight into daily anatomical changes [9]. However, low CBCT image quality in the form of scatter noise, low soft tissue contrast, artifacts, and lack of calibration procedures hamper the direct use of CBCTs for dose calculations or delineation propagation in adaptive workflows [[10], [11], [12], [13]]. Different solutions have been proposed to overcome CBCT image quality limitations, ranging from deformable image registration to artificial intelligence-driven reconstructions and synthetic CT (sCT) generation [[14], [15], [16], [17], [18]]. Industrial OART solutions that include CBCT-based sCTs are already available [19,20]. However, most clinics have not yet adopted CBCT-guided OART due to the lack of dedicated online quality control tools, and the increase in workflow time, which mainly stems from manually correcting delineations [[21], [22], [23], [24]].
In this study we introduce a software tool to generate a sCT and accompanying delineations in which target misalignments are corrected according to the CBCT anatomy, thus enabling OART workflows. More specifically, the proposed tool is intended to correct residual GTVprim misalignments with respect to the position of mediastinal lymph nodes in LALCP. The sCT generation circumvents CBCT image quality limitations by applying a deformation vector field (DVF) to the pCT and delineations. It is fully automated, and relies on clinically approved IGRT in-room registrations. In contrast with other solutions, this tool provides a faster OART workflow, given that targets are displaced rigidly and thus manual corrections of delineations are not needed.

2
Materials and methods
2.1
Smart adapt (SA)
Smart adapt (SA) is a software tool whose feasibility was tested in a prospective trial for prostate patients [25,26]. The SA tool was built in-house, without connection to any commercial solutions. Its intended use is to correct for rigid translations or rotations of targets or other relevant anatomical structures, while propagating the rest of the anatomy without deformations [26]. In this study, SA was used to generate lung sCTs and accompanying delineations in which the position of GTVprim was corrected according to the CBCT anatomy. The SA tool featured three components, which ran in an automated pipeline [27]. It used conventional IGRT in-room local-rigid CBCT-to-pCT registrations, combined registrations into a DVF, and applied the DVF to a pCT and delineations to produce a sCT and accompanying delineations. It incorporated local-rigid translations/rotations of specific anatomical regions in the sCT and delineations.
According to conventional IGRT protocol, two rigid 3D registrations were performed in-room between the patient alignment CBCT and pCT: a carina region of interest (ROI) registration (Fig. 1a), and a GTVprim mask registration (Fig. 1b). The registrations were performed subsequently, using the same CBCT. Afterwards the treatment table was shifted without rotations, according to the carina registration. Both registrations were performed automatically based on the intensity values of both scans inside the selected regions, without rotations (‘grey value’ registration in XVI 5.0.7; Elekta Oncology Systems Ltd, Crawley, UK). The mask for GTVprim registration was created during the first treatment fraction by expanding the corresponding delineation by 5 mm. Occasionally, manual mask adjustments were needed (to exclude bony anatomy).
The SA pipeline used as input a spatial registration dicom object containing the transformation matrices corresponding to each registration, to subsequently build a DVF. Given that a treatment table shift was performed during the IGRT registration protocol, the DVF does not shift the patient within the scan. The DVF created leaves the whole anatomy unaltered (in agreement with the carina-based table shift), except for GTVprim, which should shift into the correct position according to the GTVprim registration. We refer to vectors corresponding to carina registration and GTVprim registration as DVFcarina and DVFGTV, respectively. As a first step, the residual GTVprim misalignment that remained after shifting the treatment table according to the carina registration was calculated as the difference between the coordinates from the GTVprim and the carina registration (DVFGTV – DVFcarina). As a result, a uniform DVF with the vector magnitude and direction necessary to shift the tumor in the correct position (DVFtumor) was obtained, which covered the extent of the pCT, using a spacing of 2 mm in all directions.
The vectors of DVFtumor should only be applied to GTVprim and not to the rest of the anatomy. To that end, an auxiliary mask (help mask) was created based on the GTVprim delineation and applied to DVFtumor. In this way, vectors outside the help mask remained as zero, while vectors inside the help mask retained the features of DVFtumor. To avoid abrupt changes in vectors at the transition from the tumor edge to the surrounding anatomy, a Gaussian smoothing kernel was applied to the help mask. The kernel size varied in X, Y and Z axes, according to the magnitude of the vector components. This choice ensured movement coverage in the axes along the tumor shift, while minimizing deformations in the rest of axes. Fig. 2 illustrates the mask creation process: an exemplary tumor shift with a larger component in the Z axis (see Fig. 2a) led to a help mask wider in the Z direction (see Fig. 2b). The final DVF after applying the help mask is illustrated in Fig. 2c. Fig. 2, Fig. 2 illustrate a DVF for a patient.
To generate the sCT and delineations, the pCT and delineations were deformed by the DVF. Fig. 3, Fig. 3 show the difference between pCT and sCT anatomies. The resulting sCT mitigated the residual GTVprim misalignment previously observed in the pCT after patient alignment. Fig. 3c shows sCT-CBCT agreement for GTVprim and carina alignment.

2.2
Dataset: patient and CBCT selection
Eight retrospective LALCP, treated with IGRT, were used to evaluate the performance of SA. All patients gave consent to use their data for research purposes (IRBd24-216). Main patient features, such as stage, tumor lobe location or lymph nodes involved, and treatment protocols, can be found in Table S1. Thoracic pCTs were acquired in 4D, with a 600 mm field of view, a matrix size of 512x512 pixels and 3 mm slice thickness (Somatom Go.Open Pro, Siemens, Forcheim, Germany). The pCTs were reconstructed with a mid-position algorithm [28].
A selection of 20 patient alignment CBCTs were processed by SA, to evaluate the position of GTVprim in the sCTs. In-room 3D-CBCTs (Elekta XVI 5.0.7; Elekta Oncology Systems Ltd, Crawley, UK, augmented with software for motion compensation [29]) from treatment fractions in which residual target misalignments (>= 2 mm) were present were selected. The selection focused solely on evaluating SA in terms of correction of target misalignments. This dataset does not represent overall target misalignments through the course of a whole treatment for an average population. CBCTs were acquired in free breathing as part of conventional IGRT [29]. To prevent systematic errors in the GTVprim position in CBCTs, a 4D-CBCT is typically acquired on the first day of treatment and compared against the 4D-pCT.
Since the DVF deforms the immediate surroundings of GTVprim (see Fig. 2, see small rib deformation in Fig. 3), a minimum distance between GTVprim and lymph nodes was established as a patient selection requirement. This strategy prevented undesired lymph node deformations caused by their proximity to GTVprim. The minimum distance was determined by visual inspection of sCTs from patients with different target distances (0.0, 1.2, 1.7, 2.0, >2.0 cm). Lymph node deformations were found for target distances of 1.2 cm or below, thus only patients with target distances of 1.2 cm or more were included.

2.3
Evaluation
To quantify possible residual GTVprim misalignments in sCTs after SA, each sCT was registered against their corresponding CBCT, following the registration procedure described in section 2.1. Residual GTVprim misalignments were quantified via registration vector lengths between each sCT and CBCT, as well as pCT and CBCT. Median and interquartile range (IQR) of residual GTVprim misalignments were extracted.
Differences in HU between sCT and pCT, referred to as HU errors, were evaluated inside GTVprim to assess the suitability of sCTs for dose calculations inside the target, which is the region that has been rigidly displaced. HU errors inside GTVprim delineation were quantified pixel-wise for every sCT, where each sCT was linked to a CBCT. Mean absolute HU errors (MAE) were quantified per scan. To overlap the GTVprim of the pCT with that of the sCT and perform the voxel-wise HU value subtraction, the IGRT GTVprim mask registration was applied to the pCT.
The DVF was applied to all delineations linked to the pCT, resulting automatically in a new set of delineations linked to the sCT. The GTVprim delineation was rigidly translated into a new position. This process could result in small delineation volume changes, due to the re-slicing of a displaced delineation into the sCT grid. Target delineation volumes were calculated for the pCT and the sCT, and the sCT-pCT volume difference for each CBCT was computed to measure possible delineation volume changes introduced by SA.

3
Results
Residual GTVprim misalignment vector lengths between pCT and CBCT had a median value of 5.1 mm (IQR = 2.0 mm), with a maximum of 12.6 mm and a minimum of 2 mm (see Fig. 4a). In contrast, sCTs generated by SA showed reduced residual GTVprim misalignments, with a median of 0.7 mm (IQR = 0.5 mm, minimum = 0.2 mm, maximum = 1.3 mm), demonstrating the ability of SA to represent correctly the actual GTVprim position in the CBCT anatomy.
As shown in Fig. 4b, the highest mean HU error was 6 HUs (CBCT 1), the lowest mean HU error was −1.4 HU (CBCT 7). Median values were between 0 and 1 HU for all scans. IQR of HU errors varied per CBCT, with values below 10 HUs. In cases of highest HU agreement, IQRs were within ±4 HUs. Largest HU deviations were found in CBCTs 7, 8 and 9, with maximum/minimum values within ±30 HUs. MAE per scan ranged between 4 and 15 HUs.
Volume differences were negative, showing volume reduction in sCTs (see Table 1). Absolute volume differences were below 0.7 cm3, with a median of 0.2 cm3. In relative terms, the median volume change was −3.7 %. The largest relative volume differences were −13 %, linked to pCT delineations with small volumes below 2 cm3 (patient 7, CBCTs #13, 14).

4
Discussion
Due to differential motion between GTVprim and affected lymph nodes in LALCP, a compromise is needed during daily patient alignment in conventional CBCT-guided IGRT [2,3]. In this work, an in-house developed application (SA) was introduced to create sCTs and delineations, enabling an OART workflow. Residual GTVprim misalignments with respect to the position of lymph nodes were consistently corrected, HU accuracy inside GTVprim was preserved, and target delineations were propagated without deformations.
The performance of SA was evaluated by quantifying vector lengths in pCT-to-CBCT and sCT-to-CBCT GTVprim registrations. Fig. 4a shows a consistent reduction in vector lengths from pCT to sCT for all CBCTs. A median residual vector length of 0.7 mm (IQR = 0.5 mm) showed that GTVprim in sCTs was in the correct position according to the CBCT anatomy. Residual vector lengths in sCTs were in agreement with the registration accuracy (below 1 mm [3]). Error sources that contributed to the magnitude of residual vector lengths mainly stemmed from occasional manual adjustments of GTVprim mask.
Small sCT-pCT pixel-wise HU errors were found inside GTVprim, showing the ability of SA to preserve HU values inside the tumor delineation and the suitability of sCTs for dose calculations in the target region. Given that the DVF displaced solely the GTVprim, HUs were not quantified in other anatomical regions. All median HU errors were within 1 HU. In cases of highest HU agreement, IQRs were within ±4 HUs. Tissue heterogeneities in GTVprim delineations, where lung tissue was present, led to IQRs below 10 HUs. Such small HU discrepancies are not foreseen to have a relevant dose impact [[30], [31], [32], [33]].
GTVprim delineation volume differences between sCT and pCT were quantified in Table 1. The sCT systematically showed small GTVprim volume reduction (median of 0.2 cm3, below 4 %) with respect to pCTs. This phenomenon was due to the re-slicing of convex delineations (such as GTVprim) into a new sCT grid position. Approximating GTVprim as a sphere, special attention could be paid to diameter changes >1 mm. For a sphere with 2 cm of diameter (4 cm3), a diameter decrease of 1 mm would correspond to 14 % volume change. In practice, volume changes stem mostly from the first and/or last delineation slices. Visual inspection and monitoring of volume changes are recommended. Furthermore, future studies could evaluate the impact in terms of dose of the observed volumetric changes.
The main limitation of SA is the deformation in the immediate surroundings of targets, which led to constraints on the distance between GTVprim and lymph nodes. For clinical implementation, a minimum safety distance may be considered for the spinal cord or other critical organs. This work analyzed a limited number of patients, mostly with small GTVprim tumors clearly visible in CBCTs. Tumor size does not add complexity to the method. However, special attention must be paid if SA is applied to central tumors, potentially closer to lymph nodes or critical organs. The benefits of SA might be less evident for low-visibility tumors, where CBCT-to-pCT registration may be less accurate.
New strategies to robustly generate the DVF while minimizing undesired deformations, such as the use of focus ROIs, could extend the patient population compatible with SA [34,35]. In addition to GTVprim misalignment corrections, the applicability of SA can be extended to correct for the position of more structures, e.g. bone rotations. Remeijer et al., corrected for the position of pelvic lymph nodes in prostate cancer patients using a vertebrae registration [26]. Whole-patient rotations could also be corrected, with relevance for clinics equipped with translation-only treatment tables.
In comparison with deformable image registration or artificial intelligence-driven sCTs, SA has a limited capability to reproduce anatomical changes, focused only on rigid translations/rotations, with the advantage of quick and automated processing. Furthermore, it is challenging to quantify uncertainties in deformable image registration or artificial intelligence-driven sCTs, and reaching a consensus on quality assurance metrics remains difficult [19,24]. In contrast, SA is a robust and predictable method that only requires a verification CBCT acquisition and subsequent visual inspection of sCT and delineations as a quality assurance measure.
The SA pipeline has been clinically implemented as part of an OART workflow that builds on top of conventional IGRT (see Fig. 5). Currently, OART with SA is being tested prospectively for feasibility with 20 LALCP. The pipeline is automatically triggered in each treatment fraction, after performing the CBCT-to-pCT registrations. The sCT and new delineations are generated by SA in less than 2 min, without human intervention. More specifically, SA takes 15 s to process registrations, 30 s to build a DVF, and 50 s to create a sCT and delineations. The sCT and delineations directly become available for treatment plan adaptation. For scan-specific quality assurance, a verification CBCT should be acquired after plan adaptation. The extra OART workflow steps (sCT and delineations generation, plan adaptation, and verification CBCT, as shown in Fig. 5), add 15 extra minutes to conventional IGRT. A complete OART workflow takes 25–30 min. Other solutions, such as Elekta Unity (magnetic resonance linac) or Varian Ethos, take at least 34 min [36,37].
Conventional IGRT GTVprim-planning target volumes (PTV) for LALCP have a large residual inter-fraction motion component (IFMC), due to residual GTVprim misalignments [3] (see Tables S2 and S3 in the supplementary material). The mitigation of residual GTVprim misalignments by SA allowed the revision of the IFMC for OART and the suggestion of new GTVprim-PTVs for OART. The suggested PTVs, shown in Table S4, come as a result of replacing the IGRT IFMC by values typically used for lung cancer patients with a single tumor, which feature a tumor-based table correction [38] (Table S3). A smaller IFMC resulted in a total margin reduction of 2 mm in cranio-caudal direction and 1 mm in the rest of directions.
In this study, an automated tool to correct for residual target misalignments (SA) in LALCP was introduced to address one of the weakest links in OART: the propagation of targets [39]. Targets were propagated without deformations, relying on conventional in-room clinically approved CBCT-to-CT registrations. This approach eliminated the need for a physician present in the room to perform manual delineation corrections. The SA tool was evaluated with retrospective patient data in terms of GTVprim position, HUs and GTVprim delineation volume. The sCT and delineations produced by SA represent the anatomy of the day and are ready for plan adaptation, enabling the clinical implementation an OART workflow.

Declaration of competing interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: This study was supported by an institutional grant from the Dutch Cancer Society and the Ministry of Health. Our department receives royalties for CBCT guided software from Elekta AB.
Jan-Jakob Sonke reports research grants from Elekta AB.