Deformable image registration for radiation therapy: principle, methods, applications and evaluation

Abstract

Background: Deformable image registration (DIR) is increasingly used in the field of radiation therapy (RT) to account for anatomical deformations. The aims of this paper are to describe the main applications of DIR in RT and discuss current DIR evaluation methods.

Methods: Articles on DIR published from January 2000 to October 2018 were extracted from PubMed and Science Direct. Our search was restricted to articles that report data obtained from humans, were written in English, and address DIR methods for RT. A total of 207 articles were selected from among 2506 identified in the search process.

Results: At planning, DIR is used for organ delineation using atlas-based segmentation, deformation-based planning target volume definition, functional planning and magnetic resonance imaging-based dose calculation. In image-guided RT, DIR is used for contour propagation and dose calculation on per-treatment imaging. DIR is also used to determine the accumulated dose from fraction to fraction in external beam RT and brachytherapy, both for dose reporting and adaptive RT. In the case of re-irradiation, DIR can be used to estimate the cumulated dose of the two irradiations. Finally, DIR can be used to predict toxicity in voxel-wise population analysis. However, the evaluation of DIR remains an open issue, especially when dealing with complex cases such as the disappearance of matter. To quantify DIR uncertainties, most evaluation methods are limited to geometry-based metrics. Software companies have now integrated DIR tools into treatment planning systems for clinical use, such as contour propagation and fraction dose accumulation.

Conclusions: DIR is increasingly important in RT applications, from planning to toxicity prediction. DIR is routinely used to reduce the workload of contour propagation. However, its use for complex dosimetric applications must be carefully evaluated by combining quantitative and qualitative analyses.

Introduction

Deformable image registration (DIR) involves estimating the geometric transformation between two images to map them onto a common coordinate system (CCS). The process is deformable, or nonlinear, because the estimated transformation does not include only rigid transformations (i.e., translation and/or rotation) but also deformations (e.g., shrinking or stretching).

DIR has been extensively studied for radiation therapy (RT) applications, and its integration into clinical practice is currently the object of intensive research effort as shown by the report of the AAPM Task Group 132 [1]. The main drivers behind the recent popularity of DIR for RT are the increasing number of acquired images, along with the development of multimodal imaging for planning, image-guided RT (IGRT) and adaptive RT (ART) during treatment delivery, and advances in image processing. Indeed, at planning, DIR is a particularly attractive tool to assist the segmentation when using an atlas-based model and for fusion in the case of multimodal imaging. Moreover, DIR can improve the definition of the planning target volume (PTV) through the use of population analysis. During fractionated treatment, in both external beam RT (EBRT) and brachytherapy (BT), managing organ deformations remains a significant clinical challenge. These deformations are caused by multiple factors such as patient motion, breathing, weight loss, tumors and organ at risk (OAR) shrinkage or organ filling. If not corrected, these anatomical variations in intra- or inter-fractions may lead to tumor underdosage and OAR overdosage, thus increasing the risk of recurrence or toxicity [2]. In this context, DIR has been applied to monitor the deformations, and can therefore propagate the planning segmentations to per-treatment images (i.e., images acquired during the treatment delivery). DIR can then be used to estimate and monitor the cumulated dose in a deformable anatomical structure. Additionally, DIR has been used for toxicity prediction within a voxel-based analysis framework. Thus, it allows the identification of specific organ subregions that are associated with a high risk of toxicity.

These very promising applications of DIR have led many companies to include DIR tools in their commercial solutions. As these tools can be used for different purposes, the clinical applications have very different levels of maturity. From a clinician’s point of view, when considering the use of DIR for a specific application, it is important to understand its precise role, requirements and limitations, which differ markedly among applications.

This paper describes first the principle of DIR, the difficulties for DIR, the classic DIR methods used in RT and the DIR evaluation methods, in Supplementary Materials. The paper provides then a review of the main applications of DIR in RT: at planning to improve segmentation; during the treatment for IGRT, BT, ART and re-irradiation; and for toxicity prediction. It does not cover rigid registration.

Material and methods

We searched PubMed and Science Direct for articles published from January 2000 to October 2018 using the following query: (‘DIR’ OR ‘deformable image registration’ OR ‘elastic registration’ OR ‘non-rigid registration’) AND (‘radiotherapy’ OR ‘radiation therapy’). Our search was restricted to articles that report data obtained from humans, were written in English, and addressed DIR methods for EBRT and BT.

Results

The search returned a total of 2506 articles. As this paper is not intended as an exhaustive review, 207 articles were selected based on didactic considerations to present the most frequently used methods and the main clinical applications.

Deformable image registration: principle, difficulties, methods used in radiotherapy and evaluation

This chapter is provided in details as Supplementary Material.

DIR applications in RT

The following sections introduce the primary DIR applications in the RT workflow, along with their clinical benefits and limitations.

DIR at planning: atlas-based segmentation, multimodal image fusion, PTV definition, functional imaging and MRI-based dose calculation

The first step of treatment planning is the delineation of the organs of interest. Delineation is generally based on computed tomography (CT) images, although it can involve multimodal imaging. This task can be time consuming, requiring up to 2.5 h, for example, for the head and neck (H&N) region [3].

Atlas-based segmentation

To perform automatic segmentation, atlas-based methods rely on DIR between the image to be segmented and the images of an atlas, for which the associated delineations are known. The atlas consists of a single image template or several templates (multi-atlas, typically at least 15 images, depending on the tumor localization) [4] from patients with representative anatomy and organ delineations that have been validated by experts. The first step consists of identifying the most similar template image in terms of the image intensities, specific characteristics (e.g., patient gender, age, weight) or anatomical features (e.g., organ size, centroid position, orientation) [5]. The DIR transformation between the template image and the image targeted for segmentation is then estimated and applied to the template delineations, allowing the contours to be propagated onto the anatomy of the new subject. Finally, the radiation oncologist must verify these contours and correct them if necessary. Atlas-based segmentation methods are useful for organ delineation, although they are not appropriate for the segmentation of tumors, which exhibit different shape characteristics and image features for different patients. However, atlas-based segmentation can be suitable for defining the nodal tumor volume, which is defined based on its anatomical localization. Selecting the right template is crucial, because a template that is more representative of the considered image will result in a transformation that is simpler and less prone to error. To reduce the uncertainties, the selection of a unique template can be replaced by the use of multiple templates [6,7]. In the latter case, the final segmentations over the multiple propagated contours is obtained by averaging or majority vote using the simultaneous truth and performance level estimation (STAPLE) or selective and iterative method for performance level estimation (SIMPLE) methods.

In prostate cancer, the selection of multiple templates with label fusion appears to be the best approach, yielding a Dice similarity coefficient (DSC) greater than 0.8 for the prostate, bladder and rectum [7,8]. In cervical cancer, the same approach can be used to assist the selection of a treatment plan from a planning library [9]. The generated segmentations yielded a median DSC greater than 0.8, indicating the selection of the correct plan in 93% of tested cases. In H&N cancer, a standard atlas-based segmentation method resulted in OAR and nodal clinical target volume (CTV) segmentation with a mean DSC of around 0.7 [10]. Manual correction helped to improve the mean DSC to above 0.8. Overall, the application of DIR to atlas-based methods has the ability to provide segmentations of a quality comparable to expert delineations [6–10] while also reducing intra- and inter-observer delineation variability and decreasing the delineation time (e.g., up to 40% reduction for H&N [10]). Atlas-based segmentation methods are currently offered in several treatment planning systems (TPSs) [11,12] and are routinely used in clinics.

Multimodal image fusion

Morphological and functional images, when considered as diagnostic images, play a crucial role in RT planning. For example, magnetic resonance imaging (MRI) and nuclear imaging (e.g., fluorodeoxyglucose-positron emission tomography (FDG-PET), single-photon emission-computed tomography (SPECT)) are used because of their ability to improve the tumor delineation [13]. These multimodal images must therefore be fused to combine the associated information in a CCS. However, delineation could prove to be complex when anatomical deformations or body position variations occur during image acquisition, rendering ‘standard’ rigid registration inappropriate. For prostate delineation, MRI enables to clearly define the prostate contour as well as tumor volumes, whereas CT often overestimates this organ volume and does not show the tumor [14]. DIR methods have therefore been used to propagate the delineation of the prostate from MRI to CT using a free-form deformation (FFD) [15] or finite-element model (FEM) [16]. In the treatment of cervical cancer, DIR has been used to propagate the pre-BT MRI to the planning CT. The resulting propagated high-risk CTV was shown to be smaller when using MRI than in CT [17]. For H&N cancer, DIR has been used to propagate the gross tumor volume (GTV) from the diagnostic position MRI to the planning CT, thus improving the accuracy of the GTV delineation [18]. DIR has also been applied in H&N cases to propagate CT/FDG-PET or MRI/FDG-PET to the planning CT [19,20], and is more accurate than rigid registration alone in terms of the target registration error (TRE), center of mass of the tumor, and normalized cross-correlation between the PET images. Finally, in lung cancer cases, DIR has been used to deform the static CT to a 4D MRI using a hybrid DIR method to generate a 4D CT for RT planning [21].

DIR for deformation-based model PTV definition and functional planning

DIR has been used to quantify the mobility and deformation of organs within a population of patients. The deformations can then be quantified, at both the intra-patient scale (e.g., using 4D CT or early per-treatment imaging) and inter-patient scale (e.g., using planning or per-treatment imaging), thus generating a population-based deformation model. This model can then be applied to a given patient to define the PTV margins, dose coverage probability and functional imaging-based planning. In prostate cancer, the correspondence between inter- and intra-patient anatomies can be generated by DIR, allowing a statistical deformable motion model to be constructed [22]. By accumulating the dominant deformations of the model, a population-based PTV could be defined based on prostate displacements, such as in a ‘standard’ approach population-based PTV, while accounting for deformations. The same method has been applied in cases of rectal cancer [23]. The population-based PTV significantly improved sparing of OARs compared to uniform and nonuniform CTV-to-PTV margins. In cervical cancer, a DIR-based model has been generated to simulate different treatment scenarios and assess the dose coverage probability of the CTV and OAR [24]. A similar DIR-based model was used to generate a patient-specific planning library that predicts systematic deformations of the uterus [25]. A simpler use of surface-based DIR has been proposed to estimate intermediate positions of the uterus between full- and empty-bladder anatomies, allowing a planning library to be generated [26]. The same approach has been used to construct a planning library for bladder cancer [27]. Such adaptive planning library strategies, defined before RT treatment using DIR, improve the CTV coverage of numerous pelvic tumors [28–30]. In the case of lung cancer, DIR can be used to estimate the deformation vector field (DVF) between each breathing phase of the 4D CT or the exhale/inhale images, resulting in functional imaging maps [31,32]. These maps can then be used to guide the intensity-modulated RT (IMRT) dose planning and spare functional regions, especially for patients who have high-functional lung adjacent to the PTV [33]. However, the definition of these functional imaging maps is influenced by the choice of the DIR method, and should be used with caution [34].

DIR for MRI-based dose calculation

With the development of the MRI-linear accelerator (LINAC), the dose calculation based on MRI appears to be especially crucial [35–39]. Atlas-based approaches have been designed to generate, from a single MRI, a pseudo-CT (i.e., a virtual image that mimics CT densities on the MRI anatomy) from which the dose distribution can be calculated. The considered atlas consists of pairs of associated CT and MRI images, and the template MRI can be deformed toward that of the new patient. This deformation can therefore be used to propagate the electronic densities of the template CT onto the new patient MRI space, thus providing a pseudo-CT. Studies on prostate [40,41] and H&N [42] have demonstrated that pseudo-CT images exhibit low Hounsfield unit differences compared with ground-truth CT images.

DIR in IGRT: contour propagation and fraction dose estimation

DIR for contour propagation

Three-dimensional per-treatment imaging (e.g., CT, cone-beam CT (CBCT) or MRI) is routinely employed for tumor positioning under the LINAC system. Per-treatment imaging using DIR can also be exploited to characterize anatomical deformations and their impact on the fraction dose. Thus, DIR has been used to propagate delineations from the planning CT to the per-treatment imaging (Figure 1). A DIR transformation is computed between the planning CT and per-treatment image, thus providing a DVF. The DVF is then used to propagate the planning CT delineations to the per-treatment image. An expert must then validate and, if necessary, correct the delineations. For example, in prostate cancer, a non-parametric or FFD DIR method has been used to propagate the delineation from the planning CT to the per-treatment images (CT or CBCT) [43–48]. Compared to rigid registration, DIR improved the accuracy of the organ delineation, achieving an increase in DSC. The largest and most complex deformations observed in the bladder and the rectum, however, cannot usually be fully handled by these DIR methods, and thus still require manual corrections [47,49]. Specific geometric approaches based on a B-spline interpolated transformation of the previously extracted salient points have therefore been developed, and these improve the registration to more than 90% accordance with manual contours [50]. In cervical carcinoma, constrained FFD methods (using a prior deformable model or landmarks) have been used to propagate delineations between either intra-patient [51,52] or inter-patient MRIs [53]. For intra-patient DIR, DSCs of approximately 0.85 were recorded for the bladder and uterus, which is comparable to expert delineation [52]. For inter-patient DIR, large and complex deformations between inter-patient cervix-uterus shapes were poorly handled (mean DSC of 0.55 for the CTV) and required manual correction [53]. In image-guided BT, DIR has been used to propagate high-risk CTV and OAR from fraction-to-fraction MRIs, producing clinically acceptable dose uncertainties [54]. In liver cancer, a commercially available DIR method has been used to propagate abdominal organ delineations from the planning CT to the per-treatment CT with similar geometrical accuracy as inter-observer variability within 20 s per CT [55]. A biomechanical-based DIR can propagate the GTV delineated on the planning CT to the per-treatment CBCT based on a liver surface mapping [56]. This approach can be used to track the GTV during EBRT treatment for liver cancer when the per-treatment imaging contrast limits the visualization of the tumor. In H&N cancer, intra-patient DIR has been used to propagate the planning contour to the per-treatment CTs [48,57–59] and CBCTs [60–62]. Additionally, when propagating the contours from the pretreatment MRI to the end of treatment MRI, DIR provided a contour error similar to the voxel size (2 mm) and a DSC of around 0.8 [63]. Overall, DIR was consistently more accurate than rigid registration, with a mean DSC increase of 0.12–0.15 and better TRE (3 mm on average). In lung cancer and upper-abdominal malignancies, tumor volumes are manually delineated on each phase of the 4D CT [64]. By enabling GTV propagation between the 4D CT phases [65–70], DIR reduced the delineation time by a factor of 2 (from 40 min to 18 min) [65,66]. The propagated GTV delineations were similar to the manual delineations [71], with a reported DSC greater than 0.76, which is similar to the reported intra-physician variation score [65,67]. Thus, DIR can reduce the time required for the delineation process while overcoming inter-observer variability. However, the DIR uncertainties require a physician to check and correct the propagated delineations. DIR can account for anatomical variations caused by breathing, which is particularly interesting for the incoming MRI-LINAC systems, as the delineated tumor at planning can be propagated onto the per-treatment 2D MRI for the purpose of gating the treatment beam [72,73].

Figure 1. Workflow of delineation propagation from the planning CT to per-treatment image (head-and-neck). The planning delineations are propagated by means of the deformation vector field estimated by DIR. The propagated delineations are validated by the physician and corrected if needed. DIR: deformable image registration.

DIR for fraction dose calculation

If the planning contours can be effectively propagated to the per-treatment images, estimating the delivered fraction dose may become more difficult. Indeed, the planned dose distribution might not correspond to the delivered fraction dose distribution because of anatomical variations. In cases that involve large external contour and electronic density variations, the dose must be recalculated. When irradiating prostate cancer, the hypothesis of dose invariance (i.e., the dose map remains globally constant between the planning and the fraction) has been validated, except in the event of the significant appearance/disappearance of rectal gas or large external contour variations [74,75]. When irradiating locally advanced H&N cancer, weight loss and shrinking of the tumor and parotid gland mean that dose recalculation is needed. Thus, an image that represents the current anatomy with reliable electronic densities is required. DIR has therefore been used to propagate the Hounsfield unit from the planning CT to the per-treatment imaging (e.g., CBCT or megavoltage CT (MVCT)), generating a pseudo-CT that allows the fraction dose distribution to be calculated [76,77].

In terms of H&N cancer, several studies have evaluated the accuracy of DIR in generating the pseudo-CT in terms of the electronic density difference and dose calculation (i.e., dose difference and gamma index) [60,78–81]. The dose uncertainties were found to be small, and were considered ‘clinically acceptable.’ A similar approach to deform the planning CT toward the online CBCT was investigated in a cohort of five lung cancer patients. The approach was dosimetrically evaluated in a single patient to confirm the possibility of triggering a re-planning based on anatomical changes [82]. However, the inherent noise, low contrast and limited field of view (FOV) in CBCT and MVCT make DIR methods challenging. Thus, preprocessing may be necessary to compute the dose on a corrected/modified CBCT [78,79,83,84].

DIR in ART for dose accumulation

Justification and principle of DIR for fraction dose accumulation

A crucial issue in RT is estimating the cumulated dose over the fractions, either to report the delivered dose or to compare with the planning dose (dose monitoring), to trigger, for example, re-planning. Indeed, in EBRT, the dose-volume histograms (DVHs) cannot be simply aggregated when considering deformable structures [85–89]. Local anatomical variations must be accounted for in mapping the fraction doses to a CCS before summation. It is thought that DIR could be used to perform this mapping (Figure 2). With the same approach as contour propagation from planning CT to per-treatment images, the DVF between the planning CT and per-treatment image can be applied to deform the fraction dose toward the planning anatomy. The deformed fraction doses propagated into a CCS in this way can then be summed. Figure 3(A,B) illustrates the difference between the direct addition of the DVH and the use of DIR to accumulate the DVH while using a numerical phantom of the pelvis [90]. The significant difference in outcomes emphasizes the need for a DIR-based method to accumulate the doses. In BT, the direct addition of the DVHs overestimated the delivered dose compared to the more appropriate DIR approach [91–95].

Figure 2. Workflow of cumulated dose estimation by deformable image registration. Step 1: a deformable image registration is computed between the per-treatment images and the planning CT image. Step 2: the fraction doses are calculated from the per-treatment images with the same treatment parameters as the planning. Step 3: the fraction dose distributions are propagated to the planning CT by means of the resulting deformation vector fields. Step 4: the propagated dose distributions are summed to compute the cumulated dose on the planning CT. The planned dose can be compared to the estimated cumulated dose.

Figure 3. Cumulated DVH should not be estimated as the mean fraction DVH. DVH: Dose volume histogram. A pelvic numerical phantom was designed (A) to compare dose accumulation using ground truth accumulated dose (in red) and DVH averaging (in blue) in the bladder for prostate cancer irradiation. The fraction dose DVHs are in gray (normalized to the total dose). The ground truth accumulated dose was obtained by propagating the fraction doses to the planning using the reference DVF resulting from biomechanical laws (Supplementary Material). The cumulated dose appears superior to the planned dose. The mean fraction dose overestimates the dose received by the bladder.

After the DVF estimation, two approaches can be used to propagate the dose accordingly. The first is to linearly interpolate the dose on the spatial grid of the fixed image; although, this does not consider the physical properties of the tissues in terms of the dose absorption [96]. The second approach considers the tissue density when recalculating the deformed dose. Moreover, dose propagation methods should follow the principle of energy conservation [97]. Monte Carlo-based dose propagation has proven to be accurate for regions that have heterogeneous electronic densities (e.g., the lungs) [98], whereas linear interpolation-based dose propagation can be used for homogeneous regions (e.g., the pelvis).

Clinical applications of DIR for fraction dose accumulation in EBRT, BT and re-irradiation

DIR enables the accumulation of doses during EBRT and/or BT and, in the case of re-irradiation, in a large number of tumor localizations. In prostate cancer, OAR deformations mean that differences between the planning and cumulated dose have been quantified [99,100]. The mean dose difference is reportedly 7 Gy for the bladder and 2 Gy for the rectum [99]. The use of DIR has also allowed the decrease in rectal dose when using a rectal balloon to be quantified (70% of patients showed a decrease of more than 5% in the normal tissue complication probability, NTCP) [100]. Moreover, DIR can be used to accumulate the rectal dose of EBRT and high-dose-rate (HDR) BT [101]. The accumulated dose was higher when using DIR (21.3% for D_0.1cc, 6.3% for D_1cc and 3.5% for D_2cc) than that given by the direct addition of the DVH. In bladder cancer, DIR has been used to accumulate the treatment dose on the planning anatomy, allowing the comparison of several ART strategies [102]. For cervical cancer, the use of DIR to accumulate the EBRT doses did not show large differences compared to the planned dose for the target; although, large dose differences have been observed for OARs [103–105]. A recent study compared the performance of different RT and ART strategies in terms of accumulated dose for a single patient, showing good conformity with the geometric CTV-PTV coverage [106]. In HDR-BT, DIR can also be useful to accumulate the fraction doses caused by the large deformations resulting from organs filling and intracavity applicator (re)insertion [91,107–109]. Moreover, DIR has been used to deform the EBRT planning CT anatomy toward the planning CT/MRI anatomy at the time of BT [110], allowing the cumulated received dose to the rectum and bladder to be estimated [111]. However, uncertainties in the dose accumulation are linked to the DIR accuracy and the EBRT technique (i.e., steep dose gradient) [92,112,113]. Moreover, the ground-truth is unknown, thus making the evaluation of cumulated dose more difficult. The clinical benefit of DIR in this context, therefore, must be evaluated [95]. In H&N cancer, weight loss along with the shrinking of the tumor and parotid gland during treatment can lead to large differences between planned and delivered doses [114]. DIR methods estimated a mean overdose in excess of 3 Gy for at least 30% (up to 60%) of parotid glands [3,115–118]. However, a study on 133 H&N patients with MVCT image-guidance reported only a small deviation in the spinal cord delivered dose estimated by DIR in the case of large anatomical changes [119]. In lung cancer, breathing causes anatomical variations during treatment delivery, which motivates the use of DIR to accumulate the dose for each individual fraction. Using biomechanical model-based DIR for the planning 4D CT and per-treatment 4D CBCT of 10 breathing phases, the average dose difference between the accumulated and planned dose, considering the minimum dose to 0.5 cm³ of the tumor, was quantified as 0.8 Gy; although, the difference was more than 2 Gy for 30% of patients [120]. Other studies have investigated the impact of breathing on dose accumulation, suggesting large variations in the OARs [121,122] and none at the target [71].

Finally, DIR has been used in the case of re-irradiation to propagate the dose from the first planning to the re-irradiation planning, such as in H&N, brain, liver, mediastinum and lung cancer [85,87,123]. DIR appears, therefore, to be especially helpful for guiding radiation oncologists in irradiating the recurrence while avoiding increased toxicity. However, the use of DIR in this situation appears challenging because of the significant anatomical differences, especially when related to matter (dis)appearance between the two treatments.

DIR for toxicity prediction via voxel-wise population analysis

DIR can also be applied to investigate the local relationships between dose and side-effects by analyzing the dose at finer scales via voxel-wise population analysis, thus revealing local differences across individuals. Such methods are inspired by voxel-based morphometry (VBM), which typically assesses the differences between groups at each voxel and relates them to different covariates (e.g., age, gender, diagnosis and cognitive scores). As mentioned before, the toxicity studies have allowed the identification of more predictive subregions within the organs [124–126]. Mapping the dose to a CCS remains a central question: the map can be obtained via a parametric representation of the anatomy in a spherical coordinate system [124], or can be more precisely computed through tailored DIR [127]. However, voxel-based methods require different steps, as shown in Figure 4: (i) the mapping by DIR of a population of individuals onto an anatomical template (steps 1 and 2); (ii) the propagation of dose distributions according to the obtained DVFs (step 3); and (iii) a local statistical analysis of the dose–effect relationship (step 4) that allows anatomical subregions that are at high risk of toxicity to be identified.

Figure 4. Workflow of voxel-based analysis using deformable image registration for patient toxicity prediction. Step 1: feature extraction is done on the population data and anatomical template (i.e., preprocessing); step 2: a DIR method is used to compute the inter-individual deformation vector fields (DVF). Step 3: the resulting DVFs are used to propagate the 3D planning dose distributions of the population on the anatomical template coordinate system; step 4: a local statistical analysis of dose–effect relationship is performed.

In prostate cancer IMRT, DIR has enabled antero-inferior parts of the anorectum to be identified as highly predictive of rectal bleeding [125,126]. This subregion was found to receive a significantly higher dose (up to 6.8 Gy) in patients with toxicity symptoms compared with nontoxic patients. In H&N cancer, inter-patient DIR was used to compare the local dose of patients with and without acute dysphagia. Two subregions (cricopharyngeus muscle and cervical esophagus) were found to have significant differences in terms of the received dose (by more than 10 Gy) [128]. Finally, in lung cancer cases, DIR was used to perform voxel-by-voxel analysis to assess the relationship between a local received dose and lung injury [129]. For each patient, a probability model was developed to represent the risk of severe lung injury under the received dose. Moreover, the peripheral medial-basal portion of the lungs was found to receive a higher mean dose in patients with lung damage [130].

In summary, the identified anatomical subregions can be used for patient toxicity prediction and/or specific OAR definition, and must be particularly spared at the time of planning to decrease the toxicity.

Discussion

Uncertainties and perspectives of DIR in RT

DIR is validated in routine treatment for delineation at the planning and per-treatment contour propagation [1]. Its accuracy appears to be similar to the inter-observer variability in prostate and H&N localization [48,131]. However, its contribution is still limited for CT-CBCT DIR of anatomical localization showing large deformations during EBRT. Moreover, because of the limited number of dosimetric studies and the absence of published clinical studies, the use of DIR for dose accumulation is still under evaluation and cannot be used directly in clinical practice [3,132]. Indeed, the use of DIR for dose accumulation is subject to uncertainties that are linked to multiple parameters, such as the DIR algorithm’s performance, the order in which the considered images are registered (i.e., inverse consistency), the lack of contrast between images, the conservation of mass, tissue sliding and the dose mapping method (linear interpolation or Monte-Carlo). The performance of various DIR algorithms has been compared to quantify dose accumulation in several anatomical localizations [57,88,117,133–138]. Although volume-of-interest-based metrics provide some idea of dose accumulation uncertainties [135,137], the dose mapping error remains very difficult to quantify [136].

With the aim of improving the DIR accuracy, some commercial software packages now provide the ability to drive the DVF using contours [127,139] or corresponding points [140]. For this purpose, specific tools, developed in academic institutions, are imported to commercial RT TPS [12,139,141]. Therefore, DIR algorithms require further development to better meet clinical needs, such as accounting for the different imaging modalities between planning and treatment delivery [133,142], near-real-time algorithms with graphic processing unit-based frameworks [139,143–145], and anatomical properties simulated by finite element models [146–150].

Issues in the choice of the DIR evaluation method

While numerous studies have compared the performance of different DIR methods (commercial, open access or homemade) for RT applications [57,117,151–159], one of the main challenges that faces DIR evaluation is its adaptation to each considered application. For example, dose monitoring requires low point-to-point error, whereas delineation propagation generally only requires good correspondence between the organ boundaries. Moreover, DIR evaluation for dose monitoring must include dosimetric indices (e.g., mean dose, homogeneity index, dose–response indices), as any given geometric error could cause different dosimetric errors depending on the dose gradient (i.e., high dosimetric errors in cases of high dose gradients) [149]. A perfect point-to-point matching is probably an unreachable goal considering the voxel size, artifacts and realistic deformation constraints. Owing to the large number of DIR methods, standard evaluation criteria must be defined and generalized [153,160]. The development of numerical phantoms is a first step toward comparing and validating DIR methods, particularly for dose accumulation. Moreover, challenge datasets are available online, enabling the evaluation and comparison of various homemade DIR methods [161–165]. More advanced techniques, such as neural networks, could also be used to quantify DIR performance [166].

However, a number of particular situations will always defeat DIR methods. Thus, even after a thorough evaluation study, each individual registration should be evaluated. In general, qualitative and quantitative evaluation methods should be combined to improve the robustness of the evaluation process, e.g., the quantitative scores may be used to guide the qualitative analysis. Moreover, geometric evaluation should always be verified by DVF analysis to ensure that the DIR provides a physically plausible solution.

Conclusions

Although, there has been significant progress in the development of DIR in RT, further developments are required to enable its exploitation in the clinical workflow. DIR in RT is complex in terms of both methods and applications. As recently recommended by the AAPM Task Group 132 [1], DIR can be used at planning and during IGRT for autosegmentation and multimodality fusion to reduce the delineation workload, especially given the increasing number of images generated for each patient. Using DIR to assess the accumulated dose could be the answer to an unmet clinical need, namely the need to estimate the potential deviation from the planned dose. DIR also enables a comparison of different RT strategies, mainly ART, in terms of the delivered dose. Nevertheless, DIR must be used with caution, as this more complex application requires high local accuracy. Indeed, DIR evaluation must consider both geometrical and dosimetric metrics, which thus requires sophisticated dedicated tools such as numerical phantoms. Moreover, a consensus on the evaluation method and criteria is required before its potential contribution in the clinical workflow can be assessed. However, given that each considered situation is unique (e.g., patient anatomy, image noise), the evaluation must also be performed at the individual scale [167].

Disclosure statement

G. Cazoulat has a competing interest with RaySearch Laboratories (Stockholm, Sweden). All other authors have no potential conflict of interest.