The geometric and dosimetric accuracy of kilovoltage cone beam computed tomography images for adaptive treatment: a systematic review

Objectives: To provide an overview and meta-analysis of different techniques adopted to accomplish kVCBCT for dose calculation and automated segmentation. Methods: A systematic review and meta-analysis were performed on eligible studies demonstrating kVCBCT-based dose calculation and automated contouring of different tumor features. Meta-analysis of the performance was accomplished on the reported γ analysis and dice similarity coefficient (DSC) score of both collected results as three subgroups (head and neck, chest, and abdomen). Results: After the literature scrutinization (n = 1008), 52 papers were recognized for the systematic review. Nine studies of dosimtric studies and eleven studies of geometric analysis were suitable for inclusion in meta-analysis. Using kVCBCT for treatment replanning depends on a method used. Deformable Image Registration (DIR) methods yielded small dosimetric error (≤2%), γ pass rate (≥90%) and DSC (≥0.8). Hounsfield Unit (HU) override and calibration curve-based methods also achieved satisfactory yielded small dosimetric error (≤2%) and γ pass rate ((≥90%), but they are prone to error due to their sensitivity to a vendor-specific variation in kVCBCT image quality. Conclusions: Large cohorts of patients ought to be undertaken to validate methods achieving low levels of dosimetric and geometric errors. Quality guidelines should be established when reporting on kVCBCT, which include agreed metrics for reporting on the quality of corrected kVCBCT and defines protocols of new site-specific standardized imaging used when obtaining kVCBCT images for adaptive radiotherapy. Advances in knowledge: This review gives useful knowledge about methods making kVCBCT feasible for kVCBCT-based adaptive radiotherapy, simplifying patient pathway and reducing concomitant imaging dose to the patient.


INTRODUCTION
Kilovoltage Cone Beam Computed Tomography (kVCBCT) is imaging system that uses cone beam computed tomography methods to achieve tomographic imaging. kVCBCT started to be used for patient setup verification in the mid-2000 when Jaffray et al. 1 began researching on-gantry implementation for radiation therapy guidance. 1 The latest On-gantry kVCBCT systems involve a kV source and large-area flat-panel detectors attached to the linac gantry, often orthogonally to the treatment beam. The path of source-digital detector is circular arc (a half-scan) or orbit (full-scan). On-gantry kVCBCT systems commercially available are: the X-ray Volume Imager (XVI) by Elekta, On-Board Imager (OBI) by Varian Medical Systems, Vero by (BrainLAB AG, Feldkirchen, Germany & MHI, Mitsubishi Heavy Industries, Japan), and the kVision by Siemens. 2 kVCBCT is utilized to correct the daily treatment errors. Among the daily treatment errors are setup errors, errors due to the complexity of the MLC delivery, internal organ motion, weight loss, and structural deformation. [3][4][5] Mendes et al. 6 and Roeske et al. 7 showed a 10% variation in prostate volumes occurs during radiation therapy. 6,7 If an error is made, the therapeutic ratio will become worsened; therefore, cure rate will be reduced, radiation side-effects will increase, or the patient even can die from excessive radiation exposure. An important strategy for reducing this error is recognizing on a daily basis the location of a tumor relative to healthy tissue. The uses of kVCBCT attached to the Objectives: To provide an overview and meta-analysis of different techniques adopted to accomplish kVCBCT for dose calculation and automated segmentation. Methods: A systematic review and meta-analysis were performed on eligible studies demonstrating kVCBCTbased dose calculation and automated contouring of different tumor features. Meta-analysis of the performance was accomplished on the reported γ analysis and dice similarity coefficient (DSC) score of both collected results as three subgroups (head and neck, chest, and abdomen). Results: After the literature scrutinization (n = 1008), 52 papers were recognized for the systematic review. Nine studies of dosimtric studies and eleven studies of geometric analysis were suitable for inclusion in meta-analysis. Using kVCBCT for treatment replanning depends on a method used. Deformable Image Registration (DIR) methods yielded small dosimetric error (≤2%), γ pass rate (≥90%) and DSC (≥0.8). Hounsfield Unit (HU) override and calibration curve-based methods also achieved satisfactory yielded small dosimetric error (≤2%) and γ pass rate ((≥90%), but they are prone to error due to their sensitivity to a vendor-specific variation in kVCBCT image quality. Conclusions: Large cohorts of patients ought to be undertaken to validate methods achieving low levels of dosimetric and geometric errors. Quality guidelines should be established when reporting on kVCBCT, which include agreed metrics for reporting on the quality of corrected kVCBCT and defines protocols of new sitespecific standardized imaging used when obtaining kVCBCT images for adaptive radiotherapy. Advances in knowledge: This review gives useful knowledge about methods making kVCBCT feasible for kVCBCT-based adaptive radiotherapy, simplifying patient pathway and reducing concomitant imaging dose to the patient.
Although kVCBCT is employed as an image guidance technology to minimize geometric uncertainty, it is expected that the actual dose to the target and organs at risk (OARs) may differ from the estimated values due to multiple factors, including weight loss, motion during interfractions, tumor regression, and progression, as well as organ deformation. To better understand the dose-response relationship to further improve the therapeutic ratio, more attention has being paid to calculate the dose of the day to increase the precision of the actual dose delivery. The paradigm for radiotherapy has shifted from image guidance to dose guidance. On a practical level, using kVCBCT for online planning to deliver palliative radiotherapy has been proven effective, with an adequate dosimetric accuracy and planning time. 8 As such, it can be used for both verification parameters, such as patient anatomy and set-up, and online dose calculation. kVCBCT image for dose calculation enables physicists to verify the actual distribution of dose on any given treatment day. Nevertheless, kVCBCT systems are limited in their ability to perform dosimetric verification in curative radiotherapy. In other words, Hounsfield Unit (HU) values are inherently unreliable for dose calculations for many reasons. First of all, high scatter photons participate in the transmitted projections by virtue of a large irradiated volume. Secondly, the relationship between HUs and attenuation coefficient (µelec) may differ for different patients and organs of the same patient. Finally, acquisition parameters, for example the number of projections, limited gantry rotation speed, tube current and potential, cause the degradation of the image quality. 4,[8][9][10] As a result, kVCBCT images cannot be used directly to calculate dose distribution.
In cases where kVCBCT images demonstrate significant anatomical deformations, only using another standard CT imaging can provide a reliable dose distribution measurement. In fact, this procedure is labor-intensive and gives additional dosage to the patients in terms of concomitant imaging dose. 11 It has been proposed many methods for converting kVCBCT into computed tomography (CT)-equivalent representations, commonly called synthesized CT, that can be used for dose calculation, treatment planning, and adaptive radiotherapy. Also, many of these methods could be used for propagating contours on planning CT (pCT) images into CT-equivalent kVCBCT or kVCBCT for different parts of the body. 12,13 In spite of large number of scientific literatures covering this subject, an inclusive overview and meta-analysis of the geometric and dosimetric accuracy of kVCBCT images for adaptive treatment is still lacking. Therefore, there are illusion about how kVCBCT can be used for adaptive radiotherapy.
The aim of the current review hence was to provide a systematic review and meta-analysis of the geometric and dosimetric accuracy of kVCBCT for treatment replanning. By providing a set of recommendations for reporting future kVCBCT dosimetric and geometric studies, the advantages and disadvantages of the techniques reported are critically assessed by evaluating metrics for dosimetric and geometric analysis, feasibility, and reproducibility.
The process of both offline and online image review "Adaptive Radiation Therapy (ART) is a closed-loop radiation treatment process, where the treatment plan can be modified using systematic feedback of delivered dose information. It intends to improve radiation treatments by monitoring treatment variations and incorporating them into reoptimization of the treatment plan". 14 ART can be classified into Offline or Online ART depending on a particular selection of the ART timescale. Whilst Offline ART is performed by imaging the patient in time interval between fractions, Online ART is performed immediately prior the treatment fraction. The former uses workflow and tools of conventional treatment planning, which makes it requiring few specialized tools; however, interfraction anatomical variations might take place when using Offline ART that will lead to inducing of further geometric error. The latter uses integrated tools to the delivery control system. Therefore, the problem of interfraction anatomical variations can be solved when using Offline ART. Online ART also faces some challenges, for instance large intrafractional variations and lack of measurements availability of patient-specific quality assurance. 15 Description of adaptive pathways Immediately prior to a treatment fraction, kVCBCT system integrated to the linac is used to acquire kVCBCT images. These images are registered with pCT to verify the patient setup. The registered image represents the starting point for ART when at this point radiotherapy professionals decide whether the original plan needs to be modified in case of structural differences between pCT and kVCBCT. When there is identification of the clinical need for adaptation, a new physician orders and radiation therapy prescription, new simulation, new treatment planning, new localization, new imaging, new assessment, new replanning, new quality assurance and new delivery are introduced. 15 In this situation, kVCBCT images would reduce further stressors on patients with mobility issues, avoid delaying a treatment start for new patient by an additional CT simulation time slot in a busy radiation oncology center, radiation exposure for patients and cost if and only if possibility of kVCBCT-based dose calculation is performed.
kVCBCT treatment replanning methodology Several strategies have been employed for correcting kVCBCT HU for dose calculation and propagating contours from pCT images into kVCBCT. The main categories of approaches used for dose computation are deformable image registration (DIR), deep learning algorithm, calibration curve, and density override, combined and hybrid algorithm. For contours propagations, on the other hand,DIR and deep learning algorithm were used.
kVCBCT for dose computation Calculating dose with kVCBCT system is not very accurate because HU values are inherently unreliable when calculating dose. Numerous factors make HUs inherently unreliable for dose calculations. Firstly, the kVCBCT HU sensitivity increases as the object size increases since large irradiated volume contributes to a large scattering of photons in transmitted projections. Furthermore, a patient's HUs and µelec may differ depending on the organ they belong to. Finally, acquisition parameters, such as the number of projections, the rotation speed of the gantry, tube currents and potentials can degrade the image quality. 9,10,16,17 In the process of correcting kVCBCT HU for dose calculation, many correction strategies have been used as described below.

Deformable image registration
The process of image registration involves aligning images in order to connect corresponding features. A registration is achieved by modifying an image by altering its geometry and intensity through geometric operations. An image is registered by aligning features with physical locations or with computer simulations. Alternatively, the images might have been collected with the corresponding sensor at different intervals or with a variety of sensors; for instance, each sensor may be sensitive to a distinct part of the electromagnetic spectrum.
A geometric operation converts a given image I into a new image I ' by modifying the pixels' coordinates. By assuming the intensities of the image are unchanged but their positions do, this means the image function value (I ) at the original location (x, y) is transferred to the new location (x ' , y ' ) in the transformed image I ' by applying the transformation function (T) that is required to model this process. This section discusses image registration applications used in medical imaging. This covers a variety of image usage but focuses primarily on imaging used in in room radiotherapy.
In image registration, the similarity between images is evaluated using complex methods, including cues and content, to identify similarities between images. There are three methods of DIR, which are physical model methods, model-based methods and hybrid methods. While physical model methods include intensity-based methods and Biomechanical-Based Modelling, model-based methods include point-based methods, and hybrid method. 12,18,19 The next subsections deal with intensity-based methods, point-based methods, biomechanical-based modeling and hybrid methods that use one method as the first condition to another one.
Intensity-based methods: A method based on intensity values operates on the intensity values of the whole image content without removing any features first. A smooth transformation is always searched for that maximizes the measure of intensity-based similarity. The problem with these methods is that, despite the fact that they can act without user interaction, they have high computational costs and require preregistration to register because the two source images should be close enough. 15,20,20,21 Intensity-based methods can be classified into optical flow, demon's algorithm, and level-set algorithm. for more details, see (Yang et al. (2011) and references therein. 22 Optical flOw An optical flow is the apparent velocity distribution of brightness patterns in an image. An important information concerning the spatial organization of the objects observing the change rate of this arrangement can be obtained by the optical flow. It ascends from the relative motion of the observers and objects. ' Figure 1' demonstrates the concept of optical flow. 23 As can be seen in the figure (8), the point in image (a) is (x, y), and at time (t+ ẟt), that point has moved to a new location, which is (x +ẟx), (y+ ẟy), (b). Thus, it is clear the displacement of the point (x, y) can be said to be (ẟx, ẟy). If 1. Image points remain constant in brightness over time.
The brightness of objects in successive images remains constant as points move in space. I (x + ẟx, y + ẟy, t + ẟt) = I (x, y, t).
2. The spatial displacement (ẟx, ẟy) and the time step (ẟt) are assumed be small. 24 DemOns' algOrithm Matching images using Demons algorithm relies on the theory of diffusing models. Two images are matched by treating one image as a deformable grid, the other images as a semipermeable membrane through which the deformable grid diffuses through the channels created by effectors inside these membranes. Demons' algorithm is an application of Maxwell's Demons to clarify an inconsistency of thermodynamics developed in the late 19th century. Imagine a semi-permeable layer containing a set of 'demons' and separating two particles, a and b. Suppose also that this layer can distinguish between the various types of particles, and logically allow particle types a and b to diffuse to only their respective sides of the layer. Therefore, the only particles in A are a and b in B. In contradiction with thermodynamic principle number two, this results in a decrease in entropy. By distinguishing the particles, the demons generated an increased amount of entropy; therefore, the paradox solved and the total entropy system has increased. 25 In theoretical form, a group of 'demons' operates on the voxels of the fixed image to cause the voxels in the moving image to be moved in conformity with the fixed image. Demons' algorithm is used to calculate the vector field ((dr)= (dx, dy, dz)) to each pixel (r = (x, y, z)) that relates a moving image with a fixed image by using I m (r + dr) = I s (r)). There are six different versions of the demon's algorithm, but they differ in how dr(x) is computed 9,26 Where I m means the moving image intensity and I f means fixed image intensity.

level-set algOrithm
Level-set algorithm is another class of intensity-based method. Unlike optical flow, Curve/surface evolution theory is used to register two intensity images by evolving one image into another one by determining the velocity field of flow explicitly. 27 In parametric form, if there are two images, which are source image I 1 (X) and target image I 2 (X). I 1 (X) is registered to I 2 (X) by evolving the level-sets that evolve along their normal till I 1 (X) has the characteristics of I 2 (X). Evolution is represented as: Where S is the speed and ∇ is the gradient of the image. The evolution will stop when image I (X) changes from image I 1 (X) to image I 2 (X), so we must include a stopping mechanism in the rate term. Therefore, S = I 2 (X, t) -I(X). The vector field is calculated explicitly by the driving equation above. Therefore, velocity field ( V t ) is Point-based methods (spline algorithm): Spline is a model that is used to classify spatial transformations. Mathematics defines a spline as a polynomial determined piecewise. Often used in computer graphics and computer-aided design, splines are popular because of the ease with which they can be constructed, the ease with which they can be evaluated, and their ability to approximate complex shapes through the designing of the interactive curve. 28 Splines were originally used to model the surfaces of planes and ships using long strips of metal or wood by attaching different weights to them along their length to bend them. An example of using splines to represent spatial transformation is by applying two surfaces where vertical displacement links to the height above the plane. For spline registration, matching landmarks or points must be identified in the target and source images. These points are called control points at which using splines, transformed images are mapped from the control point's location in the target image into its counterpart in the source image by either approximating or interpolating displacements, which are represented as a whole a smoothly fluctuating displacement field. A spline-based mapping function is defined by either determining the control points of geometrical and structural landmarks identified in both images or using quasi-or pseudo landmarks (a regular mesh can be formed by placing control points equidistantly across an image).
The interpolation condition can be modeled as where x i and x ′ i symbolize positions of the control points in the target image and the source image, respectively (Hajnal & Hill, 2001). 29 Biomechanical-based modeling: Biomechanical-Based Modeling is a model that is used to model tissue distortions in imageguided radiotherapy. In this model, the finite element methods (FEM) is used to solve the partial differential equation (PDE) for elastic deformations. In this technique, the properties of rigid, fluid, and elastic structures are represented by a three-constituent model: 1. A triangular mesh with N knots is exploited to split the image for this purpose. 2. Anatomically-based labels are assigned to each knot based on their physical characteristics. Whilst rigid label is given to bone, CSF is given to fluid, but elastic label is given to soft tissues. 3. Rigid labels are Remained fixed, fluid and elastic labels are deformed by lessening a function of energy.
Deformations can be constrained by folding energy for fluid labels, but either a stiffness energy or a tension energy for elastic labels. This energy plays a central role in avoiding the transformation from the The fold energy is calculated by: The area of the undeformed triangle is represented by A 0 , the area of the deformed triangle is represented by A, and a threshold of energy is Υ .
The stiffness energy is calculated by: The tension energy is calculated by: . ϕ i,j or k represent nodes, ϕ 0 i, j represents the relaxed space between two nodes.
A similarity measure is used to diminish the distance between matching points to devise the registration (Edward et al., 1998).
Hybrid DIR: Hybrid methods include combining two DIR algorithms or using output of one algorithm such as deformable vector field (DVF) as the first condition to another one. 12,30 Calibration curve: There are two types of calibration curves used for dose calculation on kVCBCT, which are kVCBCT calibration curve and pCT calibration curve. kVCBCT calibration curves are derived through phantom/patient/population-specific measurements, while pCT calibration curves are derived from pCT images.
Density override: kVCBCT images are overridden with either the CT densities or HU values from the CT images. Once kVCBCT HU values are overridden, the dose is computed on the modified kVCBCT images. 31 Combined techniques: Combining the kVCBCT DIR with HU override, or calibration techniques, the kVCBCT is modified.
Deep learning algorithm: Deep learning is a subdivision of AI that belongs to the machine-learning branch.
In deep learning, neural networks are used to teach the input data a specific task to yield hierarchical demonstrations of these data automatically. Deep learning has been proposed for image processing, especially generation of syntheses CT (sCT), which belong to convolutional neural networks (CNNs) class. Throughout training process, parameters learned by this process are used to combine convolutional filters, and multiple layers of filters are used to provide the depth. Uncovering parameters of the "optimal" model normalizes the training by the search criterion specified by a loss function (L ).
In deep learning, an image-to-image translation issue is used to devise medical image synthesis by uncovering a model that maps moving image (I a ) to a base image (I b ). The most widespread CNN-based architectures for medical image synthesis are generative adversarial networks (GANs), the U-nets and cycle-consistent GAN (cycle-GAN) ' Figure 2'. GAN architecture uses two networks which are generator (G) and discriminator (D). D is trained to categories if synthetic images (I' b ) generated by trained G are actual or false to enhance G's performances. U-net architecture uses two paths, which are an encoding and a decoding path, in addition to skip connections. Paths and connections extract and restructure image characteristics, thus being taught to go from domain of I a to I b . GANs are trained with a loss that compiles (GANs) and the U-nets resulting in true images. Complimentary advantages of these given bases can be taken into account by that many of GANs' parameters can be set, and U-nets is used as a generator in the framework of GAN. Cycle-GAN is a specific derivation of GAN in which unpaired image-to-image translation can be used. two GANs, which are forward pass and backward, are trained to yield two consistency losses L C to diminish differences between I a and 'I b , and I b and I' a . In forward pass, GAN goes from moving I a to I b , but in backward pass, GAN goes from I b to I a to facilitate unpaired training. 32

METHODS
The PRISMA revision was used in this study to process and report the data. 33 The main documents adopted were '' 33 Checklist'' and ''PRISMA 2020 flow diagram''. ' Figure 3' explains the design and methodology used in the systematic review. In this systematic review and meta-analysis, phantom studies and retrospective studies of cancer patients were incorporated using the following design: treatment is radiotherapy alone using linac equipped with kVCBCT. The studies that were included are case reports, retrospective studies, and experimental studies having consequences that influence the framework of radiotherapy.

Information sources
Studies from 2005 onward that evaluated the kVCBCT feasibility for radiation treatment replanning in terms of dosimetric and geometric evaluation were located on the online databases (Google Scholar, PubMed, Scopus, EMBASE, and Science Direct). These systematic reviews are studies related to kVCBCT as tools for dose calculation and automated structures outlines. The search was taken in headings and a filter such as dosimetric evaluation. kVCBCT and contour propagation. The titles and abstracts of the included studies were not the only sources checked, but the bibliographies in the included studies were also examined. Both English and non-English language studies were searched from 2005 to 2022. 52 studies that evaluated kVCBCT images for radiation treatment replanning were included in the systematic review.

Eligibility criteria (the inclusion and exclusion criteria)
From 2005 onwards, any trial evaluating methods of evaluating kVCBCT for dose calculation and geometric evaluation was included. Only comparative studies were reviewed in this field because no randomized controlled trials were available. The inclusion criteria used the following standards: (a) evaluations in calculating the dose using kVCBCT, (b) evaluating the geometric values of automated propagation of contouring structures on kVCBCT, (c) patients' data were analyzed and compared with standard data to calculate dosimetric and geometric metrices such as dose difference, γ analysis for dosimetric outcomes and Dice Similarity Coefficient (DSC) and registration error for geometric outcomes, (d) the study should be published in a journal with high impact factor, at least (3.5). The exclusion criteria used the following standards: (a) studies of graduate or postgraduate theses, animal experiments, and scholarly reviews, (b) repeatedly published literature or similar literature, (c) studies with good quality, but they have missing data, and (d) papers deals with dental kVCBCT. Authorship of studies was not blinded during the review process since it does not add any measure of quality.
To ensure the validity of the results, the quality of the included studies was assessed according to the criteria suggested by Whiting et al. 34 A Whiting assessment, which is called Quality Assessment of Diagnostic Accuracy Studies (QUADAS), consists of fourteen statements that have been framed as questions.
Process of data collection and data extraction Information was acquired by searching the databases specified above, and the abstracts or entire papers were evaluated by three reviewers (H.A.N, G.G, and N.B) independently. having read the title and abstract of potential eligible studies, articles were assessed and included based on the inclusion and exclusion criteria. After retrieving, reading and analyzing the articles, they were reviewed according to the guidelines. 33 For instance, patients and results should accord with the review problem and object, respectively. Finally, study quality was determined according to criteria specified in (section 2.2 study quality). Basically, the entire texts of articles were read for equivocal studies before making a decision. The results were then reviewed by a third investigator if there was still disagreement. A piloted form was used to extract valuable data from the studies included, with disagreements resolved through the exchange of ideas and discussion between study participants. Data extracted from each of the included studies are as follows: (a) Study which includes first author last name and year of publication, (b) sample and sample size, (c) Journal name, (d) country, (e) software and algorithm, (d) application, and (e) metric value.

Intervention
The kVCBCT images were examined for dosimetric and geometric evaluation. While Deformable Image Registration (DIR), deep learning algorithm, calibration curve, density override, intensity scaling, artefact corrected, hybrid method, and combined method were used for dosimetric evaluation, DIR and deep-learning algorithms were used for geometric evaluation. The performance of kVCBCT for dose calculation was compared to ground truth, which represents the fan CT with geometric features of kVCBCT. The performance of kVCBCT for automated contouring was compared with contour(s) drawn on kVCBCT by experienced physicians.

Metrics
This evaluation is based on outcomes acquired from the plans prone to systematic errors. For dosimetric evaluation dose volume histogram (DVH) analysis, γ evaluation ( γ ) (2%, 2 mm) ≥ 90%, the dose differences (DD), dose similarity (DS) were used as a passing threshold. For geometric evaluation, DSC, the center of mass shift (CMS), the distance transform (DT), centroid position error (CPE), hausdorff distance (HD), distance to agreement (DTA), mean distance-to-agreement (MDA), the percentage error (PE), mean distance to conformity (MDC), normalized mutual information (NMI) and root mean squared error of the 3D canny edge (RMSEC), target registration error (TRE), the Jaccard index (JI), Error in flow endpoint (FEP), feature similarity index metric (FSIM),mean deformation difference (MDD in mm), mean mutual information difference (MID) mean absolute error (MAE), and the mean absolute differences (MAD) were used as a passing threshold.

Statistical methods
Excel Package software was used for statistical analysis. Metaanalysis was conducted using a random-effects model to evaluate the accuracy of kVCBCT for treatment replanning. It included a mean 95% confidence interval (CI), weighted average, homogeneity test (Q) and p-value. 35 The meta-analysis required studies to document the outcome of interest (e.g., γ analysis and DSC score), along with a standard deviation (SD). Statistics were used to assess the 95% CI in cases where studies reported SD. Data from studies showing suitable outcomes were aggregated for meta-analysis. Therefore, the results of the statistical evaluation were denated as a mean with ±95% CI. Subgroup analyses were not performed since there were limited data on dose calculation and automatic region of interest (ROI) contouring using kVCBCT imaging.
The γ analysis score represents a vector metric of two components that merges differences in values of local dose (∆D) and DTA. A comparative study of matrices of calculated and reference dose distribution produces a matrix of γ values. 36 In γ analysis, γ value of ≤1:0 represents the pass rate. In this meta-analysis, γ value of ≥0.9 was considered a satisfactory value for adaptive radiation therapy.
The DSC score denotes an overlap index that is used to verify segmentation images and assess reproducibility. In DSC score, DSC of 0.0 means ''no overlap'' and DSC of 1.0 means ''complete overlap''. In this meta-analysis, a DSC score of ≥0.7 was considered a satisfactory value for adaptive radiation therapy. 37 homogeneity test (Q) is ''a hypothesis test against the alternative that at least one effect size differs from the rest''. 35 It is used to check the study if it can reasonably be expressed as distributing a normal effect size. In homogeneity test, Q of >75% refers that groups are heterogeneous. Q between 0 and 40% refers that groups are with low heterogeneity. If and only if a P-value is larger than 0.05, it means there is no significant presence of publication bias.

RESULTS
In this systematic review and subsequent meta-analysis, all methods designed to make kVCBCT feasible for dose calculation and contour propagation from CT into kVCBCT were assessed against gold standards. The gold standards are the fan CT with geometric features of kVCBCT and contour(s) drawn on kVCBCT by experienced physicians for the performance of kVCBCT for dose calculation and kVCBCT for automated contouring, respectively. The main focus of the meta-analysis was to produce an overall summary of the accuracy and efficacy of kVCBCT images for radiation treatment replanning. 1008 study abstracts were scrutinized for potential inclusion; 52 papers were recognized for further estimation. The articles were evaluated using the guidelines set out by Whiting et al. 34 Overall, the included studies had satisfactory quality findings, except for patient spectrum, index scan review bias and uninterpretable scan results which scored 24%, 9%, and 24%, respectively, as shown in ' Table 1' .

Figure 4 demonstrates box and whisker plot of QUADAS quality.
Studies with a score of (yes) are more variable, especially at lower grades, and studies with a score of (not clear) vary much less than those with (No) or (yes). Therefore, it might seem that (not clear) coherence in relative grade would make predictions more dependable than the more variable (No) and (yes). Importantly, the score for (Yes) remained at the upper grade, which was about 50%, while for (No) and (Not clear) the score was 20 and 30%.
There are also significant differences between the medians of (yes), (no) and (not clear) scores. Thus, about half of the grades in (yes) have a grade of 35%, (No) have 7 %, and (Not clear) have 9%. Most of the highest-scoring 25% in (yes) are at a higher level than the highest-scoring 25% in (no) and (not clear). As a result, (yes) beats both (not) and (not clear); thus, included studies fairly satisfied quality assessment.  All articles had low bias risks. Of the 13 domains, bias was likely to be present in the subdomains of index scan execution details (29 studies) and reference standard execution details (26 studies) because of not complete details of index scan execution and the use of composite reference standards.

Literature search and study characteristics
After the computer and hand search using the strategy of search described, there were a total of 1008 studies. 800 studies were found by using computer search and 208 were found using hand search. Of total studies (1008), 863 were excluded after analyzing the title and abstract of each article because they were duplicated records, ineligible records, not in the field of interest, guidelines/reviews/books, case series/report or conference abstract, and they did not satisfy the inclusion and exclusion criteria. Therefore, 108 studies were sought for retrieval and nominated as included studies. Of these 108 studies, one study was further excluded because they were not retrieved and 38 studies were excluded due to insufficient data or other reasons. For instance, some studies did not verify the anatomical change based on the quantitative method while others did not use the standard reference for verification, but several studies used deformed CT that are not verified using a ground truth.
Finally, 52 studies were included. whilst 25 studies were included for the performance of kVCBCT for dose calculation, 27 were included for kVCBCT for automated contouring. 25 studies were conducted in European countries, 17 in the United State and 10 in Asia. Nine studies of dosimtric studies (using γ analysis as a metric) and 11 studies of geometric analysis (using DSC score as a metric) were qualified for inclusion for the meta-analysis since they come up with adequate quantitative data for the performance of the kVCBCT in radiation treatment replanning; for instance, studies have the γ values and DSC score accompanied by SD for the performance 'of kVCBCT-based treatment planning. ' Figure 5' depicts the systematic review and meta-analysis selected for this review and meta-analysis.
The review of dose calculation methods based on kVCBCT Studies that evaluated the accuracy and efficacy of CBCT for dose calculation commonly did so in terms of DVH and γ analysis for dose calculation. 25 studies were entitled for inclusion for quantitative synthesis because they were underlining the performance of kVCBCT for purposes of the dose calculation. The study characteristics and the demographics of the participants are in ' Table 2' . We reviewed the data from each study to establish metrics already described, but some publications did not have data, so re-analysis was difficult. An ideal technique would have a DVH metrics difference of 0% at a specific volume or point such as CTV, PTV, 95%, D98, and D2, and have 100% γ pass rate for dose. This level of accuracy is infeasible when kVCBCT images are used without correction. Therefore, many methods were developed to correct kVCBCT images. The types of methods used in these studies were categorized according to the model they used as follows: original kVCBCT, 38 The kVCBCT images were used to recalculate the dose for three groups of patients with cancers of head and neck (H&N), thorax, or pelvis. To calculate the dose for adaptive planning, nine studies focused specifically on the H&N while 11 studies focused specifically on the thorax, and nine studies focused specifically

BJR|Open
Original research: Checking the performance of kVCBCT geometric and dosimetric accuracy on the pelvis of 25 included studies. 25 retrospective applications of patients and ten experimental applications using phantom were conducted in these studies as displayed in ' Table 2' as well. 24 studies recalculated the dose on deformed CT or corrected kVCBCT and the ground truth was either resimulated CT on the same day of kVCBCT or planing CT if there are no significant geometric differences between CT and kVCBCT. The remaining study recalculated the dose on deformed kVCBCT and the ground truth was pCT. 39,47 Two studies calculated the dose using original kVCBCT including two implementations of H&N and one study of the pelvis. Performance assessment of these studies in terms of the validated mean dose percentage difference, 95%, D98, and D2 and Δdmax was less than 2%, but it was 5% at field edge. 38,55 15 studies calculated the dose using kVCBCT corrected with DIR or DIR combined with another method including seven implementations of H&N, seven implementations of thorax, and five implementations of the pelvis. Performance assessment of these studies in terms of the validated dose percentage difference, 95%, D98, and D2 and Δdmax respectively ranged from 0.0 to 3%, 0.0 to 9.5%, and 0 to 2.7% for H&N, thorax and pelvis [. 10,16,17,[39][40][41][46][47][48]52,56,59,60 Eight studies calculated the dose using kVCBCT corrected with one of calibration methods including six studies of H&N, three studies of the thorax, and two studies of pelvis. performance assessment of these studies in terms of the validated dose percentage difference respectively ranged from 0.0 to 10%, 0.0 to 3.9% and 0.0 to 6.7% for H&N, thorax and pelvis. 10,27,[42][43][44]50,53,57,58 Two studies calculated the dose using kVCBCT corrected with density override including three studies of H&N, two studies of thorax and two studies of pelvis. Performance assessment of these studies in terms of the validated dose percentage difference respectively ranged from 0.0 to 3.2%, 0.0 to 7% and 0.0 to 1.5% for H&N, thorax and pelvis. 9,10,57 One study calculated the dose using kVCBCT corrected with AI including one study of pelvis. Performance assessment of these studies in terms of the validated dose percentage difference respectively ranged from 0.0 to 3.7% for pelvis. 54 The dose using kVCBCT corrected with artefact corrected, including one study of H&N, one study of thorax and one study of pelvis. Performance assessment of these studies in terms of the validated dose percentage difference respectively ranged from 0.0 to 2% for entire studies. 56 Four studies calculated the dose using kVCBCT corrected with intensity scaling, artefact corrected and hybrid, but they documented rather than percentage dose difference. Therefore, they were also included for qualitative analysis as in ' Table 2 or analyzed using meta-analysis. 41,45,51,59 Overall mean difference from the gold standard dose using kVCBCT depends on the type of technique. DIR and intensity override beat others techniques with mean dose difference and SD 1% and − + 1% for both H&N and pelvis, yet calibration curve-based method beat others with mean dose difference and SD 1% and − + 1% for thorax, respectively. In the case of calibration curve-based method, the mean difference and SD from the gold standard dose were 2% and − + 2% for H&N as well as pelvis, respectively. Also, mean differences and SD up to 2% and − + 2% in the dose for DIR was observed in thorax. Artificial intelligence and intensity override achieved the highest deference in dose. While Artificial intelligence had 3.1% − + 0 dose difference for pelvis, intensity override had 3.3 − + 3.1%. ' Figure 6' indicates the subgrouping of the studies based on the technique used.

Meta-analysis of the included studies
The collected meta-analysis comprised 36 applications of kVCBCT for dose calculation, expressed in nine individual studies of included studies. Evaluation of the studies included in Meta-analysis performed in terms of γ analysis. Hence, metaanalysis revealed an overall γ pass rates of 0.98 (95% CI:0.9799-0.9801), and Heterogeneity test (Q) have 1.88 Q-value and 1 P-value, indicating that there is no significant inhomogeneity ' Figure 7' .
The review of the segmentation of organs in cone beam CT Studies that evaluated the accuracy and efficacy of kVCBCT for contour propagation did so in terms of DSC, CMS, DT, CoM, HD, RAVD, MDA, PE, and CPE for contour propagation. 27 studies were included because they were underlining the performance of kVCBCT for purposes of automated contouring ' Table 3' . ' Table 3' provides study characteristics and participant demographics based on analysis of the full-texts of these 27 included studies in the systematic review. Four groups of patients with cancer of the H&N, thorax, pelvis, or abdomen were recontoured using kVCBCT images. To automatically recontour the structures for treatment replanning, 15 studies focused on the H&N while 12 studies focused on the thorax, and seven studies focused specifically on the pelvis, but two studies dealt with the abdomen. 25 retrospective studies of patients and two experimental studies using phantom were expressed in 27 included studies as displayed in ' Table 3' as well. 26 studies propagated the contour from CT to kVCBCT and the ground truth was contour drawn on CT by expert. The one study propagated the contour from kVCBCT to CT and the ground truth was contour drawn on CT by expert. 75 Two techniques were used to contour structures on kVCBCT, including DIR and AI. While the former was used to contour the structures for twenty-two studies in which 13 studies involved H&N, ten studies involved pelvis, six studies involved thorax and two studies involved abdomen, the latter was used to contour the structures for five studies in which three studies involved H&N and two studies involved pelvis. An ideal tool would produce results as good as in the case of registering the same images of the same modality (CT or kVCBCT). Hence, the average registration error between manually contoured ROIsurfaces and automatically contoured ROI-surfaces should be 2.0 mm±5 mm, 2 mm and 3 mm in head-and-neck, the thorax and pelvis cancer cases, respectively, and DSC of (≥ 70%). 35,80,81 Most of registration errors were estimated based on the difference between calculated motion field and known motion field, manually drawn contour and manually drawn contour, center of masses, the distances to the closest point, or edges of deformed images and reference images. CMS, DT, CoM, HD, RAVD, MDA, PE, CPE and RMSEC were considered as equavalent quantity to each another, and lower values of them indicate a better agreement between the corresponding images.
Twenty-two studies used for automatic region-of-interest delineation using DIR including 12 studies of H&N, seven studies of thorax, ten studies of pelvis and two studies of abdomen. The included studies compared automatic region-of-interest contour with manual region-of-interest contour for the treatment target and the critical organs at risk. For instance, this ROI includes tumor, parotid glands and spinal cord in H&N cancer cases, vertebrae, external contour and sternocleidomastoid muscles, and tumor, bladder and rectum in pelvis cancer cases. Performance assessment of these studies in terms of the registration error ranged from 0.59 ± 0.07 to 3 ± 1.4 mm, 0.1 ± 0 to 7.9 ± 2.2 to 0.1 and 1.18 ± 0.3 to 12 mm for H&N, thorax, pelvis and abdomen, respectively. Furthermore, five studies used for automatic region-of-interest delineation using AI including three studies of H&N, two studies pelvises. According to the performance assessment of these studies, the registration errors ranged from 0.175 to 4.2 ± 1.3 mm for H&N. Seven studies were used for automatic region-of-interest delineation using DIR or AI, but they documented rather than registration error. Therefore, they were also recorded for qualitative analysis as in ' Table 3 or analyzed using meta-analysis. 12,13,62,73,82,83 ' Figure 8' indicates the subgrouping of the studies based on the technique used. Overall mean difference from the gold standard of DSC using kVCBCT depends on the type of technique. DIR and AL scored the same DSC of 0.82 ± 80.0 for H&N, yet DIR beat AI with mean DSC score ± SD of 0.8 ± 1.5 for pelvis. DSC ± SD of AI was on the other hand 0.65 ± 0.21 for pelvis. No study of AI was included for thorax. Not only DSC was used to evaluate accuracy of automatic ROI outlining, but registration error was also used in the included study as a metric. Included studies of DIR were used to automatically propagate contour from CT to kVCBCT and vice versa for H&N, pelvis and thorax, but only AI to automatically propagate contour from CT to kVCBCT in H&N satisfied our criteria. Importantly, AI beat DIR with mean DSC score ± SD of DIR v AL (2.44 ± 2.16 mm v 2.71 ± 1.35 mm) for H&N. Mean registration error of DIR for pelvis and lung were 1.79 ± 2.07 mm and 1.32 ± 0.75 mm, respectively.

Meta-analysis of the included studies
In the combined meta-analysis, there were 11 individual articles that presented 25 applications of the segmentation of organs in kVCBCT. The meta-analysis included studies that were evaluated in terms of their DSC. The meta-analysis then indicated a DSC of 0.93 (94.5% CI: 0.9219-0.9301), while the heterogeneity test (Q) showed 24.9 Q-values and 0.41 P-values, indicating a lack of significant inhomogeneity among the results, ' Figure 9' . The results of individual kVCBCT studies are also shown in ' Figure 9' . Performance estimation of these studies in terms of the validated DSC rates respectively ranged from 70% (95% CI:65.8-74.2%) to 94% (95% CI: 84.7-100.3%) for H&N, 78% (95% CI:63.1-92.9%) for thorax and 66% (95% CI:63.3-687%) to 97% (95% CI:97-97%) for pelvis, but this depends on the technique used.

Publication bias
Funnel plots were schemed using mean γ pass rate and DSC against standard errors for kVCBCT-based dose calculation and kVCBCT-based recontouring to assess potential publication bias, respectively. While applications used in the funnel plot were 36 for kVCBCT-based dose calculation, they were 25 for kVCBCTbased recontouring that were meta-analyzed ' Figure 10' . An asymmetric funnel plot signaled a bias in publication amongst the included studies. Both kVCBCT-based dose calculation and kVCBCT-based recontouring exhibited significant publication bias. The under curve area of the pseudo-95% CI and pseudo-99% CI did not encompass all studies, which supports the notion of potential publication bias. 84

DISCUSSION
kVCBCT has been utilized to compute dose and propagate contours in large numbers of studies examined in this review. H&N, thorax, and prostate cancers have been the subjects of most studies. As a consequence, many studies have suggested that kVCBCT can produce results that mimic those seen in the case of pCT images. However, the comparison has usually not been rigorous. In order to examine this, a meta-analysis of studies on kVCBCT performance in a dose calculation and automated segmentation was necessary.
Currently, resimulated CT is used for dose recalculation and segmentation of tumor lesions for adaptive radiotherapy. Adaptive radiotherapy should be considered in the first instance if it has benefit and if it does no harm in the second instance. However, using resimulated CT is associated with many Figure 7. Forest plot of studies that assessed kVCBCT for dose calculation accuracy. For dose calculation, the results from the kVCBCT are centered around a γ pass rate of 0.98 with a 95% CI ranging from 97.799 to 98.01%. Legend: CI = confidence interval; calibration curve (CC); site-specific calibration curve (SSCC); density override (DO); patient-specific calibration (PSC) and populatio-based CBCT intensity (CBCTpop).
Mean COM of oCBCT-to-rCT was (4.4 ± 2.2 mm) & eCBCT-to-rCT was (2.8 ± 1.9 mm). The difference between these points on deformed and fixed image The average difference for hybrid was 1.5 mm. (  challenges. Firstly, based on the anatomy shown on CT, patients' physical dose during prostate radiotherapy has been calculated less accurately using dosimetry calculations when compared with kVCBCT. This is because in comparison to true CT images, kVCBCT images have a slightly different anatomic distribution. 85 Secondly, anatomical changes occur commonly, which dosimetrically requires a new resimulated CT and replanning practice for treatment plan adaptation. This procedure is labor intensive, causes an additional dose to the patient and increases the cost of patient treatment. A pseudo-CT image will only inherit the anatomical information of kVCBCT as long as it has the same anatomy as kVCBCT, and another benefit is the cost saving of not having another CT scan for resimulation. Thus, by replacing current methods (CT and manual segmentation method) with sCT and an automated computer-aided system, improvement of image quality and subsequently treatment replanning can be achieved with small amount of computation time.
This review established that performance of kVCBCT depends on different methods, specific correction method itself and region of interest for dose calculation. For instance, the dose differences between dCT and pCT in H&N were 0.54%, 0.7 and 2.7% for demons, fast demons, and accelerated demons' algorithms of DIR, but these differences were 0.3%, 1%, 3.9 and 5.4% for PSCC, CC, SSCC and CC, respectively. Similarly, the dose differences between dCT and pCT in pelvis were > 2% , ≤ 2%, and ≤ 2% for B-Spline, modified Demons and B-Spline algorithms of DIR, but they were ≤ 1.3% for density override and ≤ 0.9 for combined method (DIR+override) while they were 0.6%, 1.2% or 2.3 for CC, respectively. Likewise, the differences between dCT and pCT in thorax were 5.55% and ≤ 2% or ≤ 5% for Demons and B-Spline algorithms of DIR, but they were 3.9%, 5.4, 1.9 or 2.7% and 2.4 for SSCC, CC, Stoichio-metric calibration and kVCBCT r . Importantly, in all included studies recording γ pass rates, pass rates for dose recalculation on corrected images were higher than 90% for all methods, except kVCBCT LUT and kVCBCT Pop that γ pass rates were 85% and 89 ± 8.3. According to this review and meta-analysis of promising results (overall γ pass rate of 98%; 95% CI: 97.79-98.01%), artefact correction methods mentioned above are feasible for dose calculation, but there are many variables encompassed that prevent achieving accepted dose distribution, yet these techniques have a potential capacity to improve the kVCBCT image quality.
On the other hand, this literature also determined that using kVCBCT for automatic contouring depends on propagating method itself, different methods applied and the region of interest for automated contouring. According to registration error mentioned in (Section 3.3), DIR and AI can produce results as good as in the case of registering the same images of the same modality (CT or kVCBCT), registration error of 2.0 mm±5 mm for H&N, but it cannot for pelvis and thorax because they may produce less than 70% DSC and more than 2-3 mm registration error, but the accuracy of contouring depends on the method used. For instance, values of registration error of hybrid registration algorithm and evolution registration algorithm were 0.1 ± 0.6 mm and 0.94-0.3 mm for thorax and pelvis, respectively. Furthermore, regarding the meta-analysis, most of these  results suggest that kVCBCT can be used for the segmentation of organs in kVCBCT because volume matching satisfies the recommended value (≥ 70%) for adaptive radiation therapy usages. 37 The values in ' Figure 9' are all higher than 70%, except two applications of studies by (Brion et al. 19  sCT images from kVCBCT images can be generated using a number of methods. Using AI has been limited to a small number of included studies. While one study satisfied our criteria was used for dosimetric evaluation, five studies were used for geometric evaluation. Generally speaking, AI algorithms, especially CycleGAN, U-Net GAN and FCN-GAN algorithms, in pelvis were demonstrated that they gave clinically unacceptable dosimetric results, reporting dose differences typically between 3 and 4%. Unsatisfactory results arise from the generator fails to merge deep-layer features and superficial-layer features because there is no a residual network shortcut connection contained in convolutional layers of convolutional neural network-based generator and unavailability of paired data following registration. 54 On the other hand, for geometric evaluation AI algorithms in H&N demonstrated that they gave clinically tolerable geometric results, recording DSC scores of 0.83 ± 0.06, 0.83 ± 0.09 and 0.92 for U-NET, CycleGAN and DCIGNS, respectively. Nonetheless, AI algorithms in pelvis demonstrated they were heterogeneous. For example, as with U-NET in H&N, one study gave DSC score of 0.79, but another one gave 0.63 in pelvis. 16,19,61,68,83 This can be attributed to the fact correlation between existence or the non-existence of a given body structure and intensities patterns is taught while prior information of anatomy are unavailable. For example, there is no encouragement to keep away from giving a nil guess for the structure, even although every patient possesses one, when a non-adversarial way is applied to train a network. 19 However, AI techniques are associated with many benefits in clinical radiotherapy, one of them is resolving the issue of difficult direct modification of an adaptive radiation therapy plan in addition to poor quality of kVCBCT imaging of soft tissues.  10 One of the major disadvantages of override-based is the techniques show to produce results with low level of image quality, particularly when step of override is automated. As a result, automatic segmentation of different tissue classes produces results with unacceptable outcomes. It is likely that combined methods are used in the subsequent correction steps to resolve this issue, such as DIR with override-based technique for dose calculation, and then different tissue classes are robotically segmented using DIR. 41 Calibration curve-based strategies have been showed up to provide clinically heterogeneous dosimetric outcomes, which typically depend on the method used and ROI. For instance, Dose differences are typically lower than 2% have been reported when patient-specific calibration or HU-ED calibration, stoichiometric calibration or HU-ED calibration and standard calibration curves are used for H&N, thorax, and pelvis, respectively. Hence, creating accurate reference images for dose guidance has proven feasible in these ways. 10,42,50,57 In contrast, dose differences of more than 2% were reported when using standard calibration curves and HU-ED calibrations for H & N and pelvis, respectively. 10,38,43,53,57,58 One of the main drawbacks of calibration curve-based method is that this technique has been shown to produce low quality results. Thus, automatic segmentation of various tissue classes produces unacceptable results. Subsequently, correction steps to solve this problem, such as DIR with calibration curve technique for dose calculation are applied, and then classes of different structures are automatically segmented using DIR. 52 It must be justified why any metric is used. However, the gold standard adopted should be γ analysis with clinical and statistical parameters to avoid challenges and shortcomings in dose calculation. 86 Included studies in this review often adopted DVH and γ analysis for methodologies of dose comparison for treatment plan verification. Although some included studies gave critical parameters of these matrices such as reference dose, threshold, DTA and, type of γ analysis (local or global) and dimension (2D or 3D), but they were heterogenous in several studies while other studies missed these parameters. This may be a reason of publication bias within this field of study in addition to no awareness in the publication of unwell performing artefact correction or propagation methods. This problem has to be solved by using γ analysis with consistent reporting of clinical and statistical parameters and DVH parameters for ROI in patients.
The current meta-analysis of kVCBCT for dose recalculation and automatic segmentation has a number of methodological shortcomings. Firstly, currently, a common ground truth does not exist for reference doses. Usually, authors use CT as ground truth to minimize the contribution of geometrical changes, but CT images have a slightly different anatomic distribution and suffer from non-constant image quality due to variability of image acquisition parameters. 85 Secondly, a variety of variables, such as the TPS algorithm, the tumor position, the acquisition of images, etc., affect the dose difference evaluation. Thirdly, there was considerable heterogeneity among analyses, likely because of technical differences between correction methods for segmentation. Finally, missing data resulted in excluding of several studies.
Studies have been put on the individual cases rather than population so far. Most of these studies were to evaluate the capability of image correction algorithms and models. Subsequently, many studies used few numbers of patients to test their approaches (less than ten patients). Therefore, it is noticeable that patients with untypical anatomy is inconceivable to become sufficiently tested and clinical potentiality is scarcely illustrated although this type of studies may possibly display prototype. In conclusion, a large patient cohort of clinical studies to evaluate kVCBCTbased treatment replanning is undoubtedly required.
Although kVCBCT performing kVCBCT -based dose calculation and automated segmentation have shown fairly promising results (overall γ pass rate score of 0.93; 95% CI: 0.9299-0.9399 and overall DSC value of 0.98; 95% CI: 0.9799-0.9801), this methodology is still not widely accepted and employed in daily clinical practice. This may be explained by many reasons. Firstly, standardized procedures to accurately use these correction and segmentation systems are not available. The current standards of adaptive radiotherapy and a state-of-the-art system that automatically propagates contours and scan from CT to kVCBCT are substantially different, which subsequently obstructs addition of correction methods. Secondly, the new innovation is to use kVCBCT for dose calculation and automated segmentation, but the traditional purpose for the kVCBCT use varies; where radiographers mainly use this technology for setup by comparing the pCT(or reference images) to CBCT before treatment delivery to check the positioning of the patient. This ensures reproducibility of patient set up during the course of radiotherapy treatment. Furthermore, implementation of kVCBCT methods in radiation oncology requires interaction between clinician and computer equipped with high sophisticated technology. Subsequently, trained observers are needed to supervise framework of dose calculation and automated segmentations. Moreover, there are different artefact correction and segmentation methodologies with different details that have a different effect on the consequences in applications of the same dataset. For instance, in RayStation TPS, there are intensity-based algorithm and hybrid (biomechanical model+intensity-based algorithm), but in DIRART, the top registration algorithms are optical flow and modified demons. 12,13 kVCBCT can back the adaptive radiotherapy up without the need of resimulated CT. Thus, the future application of kVCBCT in the treatment replanning is of great clinical consequence. More remarkably, sCT is highly likely to form the basis of further cutting-edge analyses to illuminate reliable and meaningful relations between features-based kVCBCT imaging and survival rate, for example kVCBCT-based dose accumulation and Prediction of radiation response of tumors using kVCBCT-based radiomics.

CONCLUSION
A systematic review and meta-analysis of various studies using kVCBCT for treatment replanning were performed to identify how kVCBCT can be used for kVCBCT-based treatment replanning performance. DIR-based method has been recognized as the most clinically valuable. This is because it can correct kVCBCT for dose calculation and contour propagation with high level of accuracy. Most of studies used in-house specific techniques, yet radiotherapy community starts becoming interested. Therefore, commercial techniques have been developed and became available. Since there is growing appeal of kVCBCT-based adaptive radiotherapy, large cohorts of patients have to be undertaken to authenticate methods within a kVCBCT-based adaptive radiotherapy workflow. Since hybrid registration method gives good outcomes of contour propagation in challenging cases, namely pelvis, these cohorts studies should focus on using hybrid registration method for different spectrum of patients with H&N, thorax, and pelvis cancers. 12,18,75 Importantly, they should focus on both dose calculation and contour propagation. Additionally, small numbers of included studies dealt with AI; therefore, further studies focusing on comparing AI to hybrid registration method-based DIR in terms of both dose calculation and contour propagation using AI for different spectrum of patients are also required. Quality guidelines should be followed when reporting on kVCBCT, which includes agreed metrics for reporting on the quality of corrected kVCBCT and validation on a set of external tests. Also, recovering quality of kVCBCT image can enhance dose calculation and segmentation results, so it is recommended to define protocols of new site-specific standardized imaging that can be used when obtaining kVCBCT images for treatment replanning.