Search

This paper presents a comprehensive extension of the Scale Invariant Feature Transform (SIFT), originally introduced in 2D, to volumetric images. While tackling the significant computational efforts required by such multiscale processing of large data volumes, our implementation addresses two important mathematical issues related to the 2D-to-3D extension. It includes efficient steps to filter out extracted point candidates that have low contrast or are poorly localized along edges or ridges. In addition, it achieves, for the first time, full 3D orientation invariance of the descriptors, which is essential for 3D feature matching. An application of this technique is demonstrated to the feature-based automated registration and segmentation of clinical datasets in the context of radiation therapy.

Purpose: To characterize the motion of the radiation isocenter, over time, as a function of gantry rotation for multiple linear accelerators (linacs). Two semi-automated image-based quality control (QC) test workflows were designed to achieve this goal. Methods: The full QC-test workflow for motion characterization consisted of acquiring 16 megavoltage images at 8 gantry angles of a ball-bearing suspended off the end of the couch. Performance constancy was assessed using a shortened QC-test workflow which consisted of imaging a cube phantom placed on the couch (5 images at 4 gantry angles). Both workflows use an image processing algorithm to determine the field center and phantom position on each image and computed radiation isocenter motion as a function of gantry angle. Motion was characterized for 9 linacs of same model and performance monitored for 2 months on 3 linacs. Results: The maximum isocenter motion determined with the full-workflow for 9 linacs was within 0.38–0.79 mm. The shortened-workflow usually agreed within 0.1 mm with the full-workflow and the time required for these methods was about 4 and 15 min, respectively. For all linacs, the isocenter motion perpendicular to the gantry rotation plane followed a consistent pattern with maximum amplitude of 0.36–0.59 mm. In the gantry rotation plane, the variation among linacs was higher and the beam axis described a circle of up to 0.6 mm radius around the gantry axis of rotation (2 linacs). The radiation isocenter motion was stable as a function of time for the monitored linacs and was within ±0.1 mm of the average. Conclusion: Radiation isocenter motion parallel and perpendicular to the gantry rotation plane was characterized. In the gantry rotation plane, beam spot positioning adjustment might be used to reduce the observed radiation isocenter motion. A shortened-workflow was designed and enables performance monitoring over time.

A theoretical method for identifying optimal X-ray imaging geometry is presented and applied to cone-beam CT (CBCT) using a flat-panel imager (FPI). The method includes the effects of X-ray source distribution, imaging task, X-ray scatter, and detective quantum efficiency (DQE) on the ICRU-recommended figure of merit for image quality-detectability index. Cascaded systems analysis for FPIs is extended to incorporate the effects of X-ray scatter directly into the DQE. For the FPI-CBCT prototype being implemented for image-guided radiotherapy of the prostate, optimal magnification is /spl sim/1.4-1.6, depending primarily on the magnitude of the X-ray scatter fraction.

Purpose: To quantify inter‐fraction variation in upper‐spine (C1–C3) positioning when using a novel vacuum bite‐block relocatable head‐frame (RHF), that was originally designed for use in intra‐cranial stereotactic radiotherapy (SRT). Methods and Materials: The RHF (Extend™) was designed for multi‐session treatments on Perfexion™. As part of an ethics‐approved study, the RHF was used to set up and immobilize patients undergoing linac‐based intra‐cranial SRT. For each fraction, cone‐beam CT(CBCT) was used to guide the skull to the treatment room isocentre based on anatomical 3D image‐matching. In this study, the same CBCT data were analyzed to quantify the relative displacement of the upper‐spine to an ideally well‐aligned skull as follows. For each patient, each daily CBCTimage set was first co‐registered to the planning CTimages based on bony anatomy in the skull. From this reference position, the mean difference and standard deviation between planning and daily treatment was determined for each of the geometrical centroid of the C2 process, as well as the spinal canal at level of C3. Results: Data from three patients were analyzed (total of 80 fractions). The mean relative displacement of the C2 process after skull alignment was 0.9±0.78 (range: −1.84–2.7) and 1.2±0.57 (range: −1.15–3.2) mm in the lateral and anterior‐posterior directions, respectively. The mean displacement at the level of C3 was 0.35±2.04 (range: −4.1–6.5) and 1.59±1.2 (range: −4.2–4.6) mm in the lateral and anterior‐posterior directions, respectively. Conclusion:_Relative to an ideal skull‐based alignment as determined with 3D‐CBCT, systematic differences in upper‐spine position were observed, as indicated by the non‐zero mean deviations. Furthermore, a larger variability in daily positioning of C3 compared to C1/C2 was observed, as indicated by the larger standard deviations. These results indicate that applying translational corrections intended for skull localization may be insufficient for localization of the upper‐spine when using the RHF.

Purpose: To explore the possibility of separately quantifying the leakage and scatter components of the peripheral radiationdose resulting from Leksell GammaKnifeR Perfexion™ radiotherapy. Methods and Materials: In this study, peripheral dose was defined as the dose arising at points outside the paths of the primary beams. Consider a spherical skull phantom centered at Leksell coordinates (100,100,100). Peripheral dose may arise from: (1) Inphantom scatteredradiation originating from primary beams; (2) Radiation first scattered within the collimator structure (3) Photons undergoing multiple scattering and exiting through adjacent collimator openings and (4) Leakage radiation that passes directly through the collimator body. Contributions (1) and (2) were separately quantified via Monte‐Carlo simulations. Types (3) and (4) were measured using EBTII Gafchromic film inserts and placing the sources in“beam‐off” position (i.e. between collimator openings). Additionally, the total peripheral dose arising from a 16‐mm shot was measured using film. The combination of simulated type_1 and type_2 doses and measured type_3 and type_4 were compared to the measured total dose at 40 mm superior and inferior to the isocentre. Results: At the isocentre of the 16‐mm shot, the dose‐rate determined by Monte Carlo simulations agreed with measured dose‐rates to 0.8%, thereby establishing consistency between measurement and simulation. The simulated sum of Type_1 and Type_2 contributions at 40 mm inferior and superior to the isocentre was 0.40% and 0.35% of the isocentre dose‐rate, respectively. Of those totals, 87% was Type_1 a 13% came from Type 2. The measured total peripheral dose‐rate was 0.40% and 0.38% at 40 mm inferiorly and superiorly, respectively. The measured Type_3 and Type_4 dose‐rates were less than 0.01% of the isocentre dose‐rate in either direction. Conclusion: The results of this study indicate that peripheral dose arising from a single 16‐mm shot arises primarily from Type_1, with

Purpose: Investigate the efficacy of three novel geometric and dosimetric on‐line strategies for adaptive radiotherapy in terms of workload and compliance with original treatment plan intent. Method and Materials: The three proposed on‐line strategies are: Selective setup corrections (S1), Combined dose‐per‐fraction adaptation and setup corrections (S2) and Optimal dose‐per‐fraction adaptation (S3). In S1, setup corrections will only be performed if the displacement is “unfavorable” in terms of a radiobiological score based on equivalent uniform dose. The score function encodes a clinical preference for maintaining dose to the target, the critical organs or an arbitrary combination. S2 combines the selective setup corrections of S1 with a change in daily dose‐per‐fraction. In S3, only dose‐per‐fraction is a variable and is determined using a rigorous constraint stochastic optimization procedure. In a retrospective study the efficacy of these strategies was investigated. They were applied to a displacement dataset of 5 prostate patients. Results: For S1, maintaining the planned dose to the prostate is costly and requires many setup corrections. In contrast, a rectum‐weighted score requires far less interventions and decreases the dose to the rectum on the order of 1.5 Gy, while approximately maintaining the dose to the prostate. Strategy S2 allows a considerable boost to the prostate dose for a preference to maintain the dose to the rectum. However, in this case, the bladder may exceed the dose limits. An equal‐weighted preference is able to moderately boost the prostate dose while decreasing rectal dose and only moderately increasing bladder dose. S3 showed the biggest potential, however, it is computationally more involved. Conclusion: Flexible correction strategies can be designed using information about treatment anatomy and planned dose distribution. Such radiobiologically motivated strategies can be made highly compliant with initial treatment plan intent while often significantly reducing the number of interventions.

Purpose: To describe and assess feasibility of a novel MR-to-Linac isocenter calibration tool for magnetic resonance guided radiotherapy (MRgRT™). Methods: The MRgRT system co-developed by IMRIS and Varian employs a movable MR system, which travels into a Linac vault. To support MR-based couch corrections based on MR-to-MR registration, we developed a software tool to calibrate the MR-to-Linac coordinate system transformation and apply it to MR images acquired for treatment guidance. To assess feasibility, we quantified repeatability of the movable MRI system at MR isocenter by securing an ACR phantom to a fixed IMRIS interventional table, and moving the MR system between the two adjacent rooms simulate normal function. We performed 18 cycles over two days in which we acquired 3D MRI datasets with 1 mm isotropic resolution. We assessed position variability by registering the MR volume for each cycle to the first volume, and tabulating transform parameters. Separately, we investigated if a calibrated position encoder on the magnet mover could minimize the z-direction shift variability. Results: Greatest isocenter variability was observed in the z direction, with mean shift of 0.21 +/− 0.12 mm over both days. The greatest rotational variability was observed about the z-axis, with mean rotation of 0.8 +/− 0.8 mrad over both days. Maximum shift was 0.51 mm in the z direction, and maximum rotation was 1.7 mrad about the z-axis. In the separate repeatability experiment we found that the calibrated position encoder reduced the mean z-shift variability from 0.59 +/− 0.35 mm to 0.17 +/− 0.13 mm. Conclusion: We showed that the movable MRI system provides a consistent within-day and between-day isocenter location, and z-shift variability can be minimized with a position encoder. This work demonstrates the feasibility of using an MR-to-Linac isocenter calibration tool for the MRgRT system. Jeff Winter, Michael Westmore and Meir Dahan are all employees of IMRIS, the company co-developing the MRgRT system. Funding for the MRgRT project at Princess Margaret Hospital was provided by the Canadian Foundation for Innovation.

Purpose: An MRI on rails is being installed at our institution to provide MR‐guided brachytherapy and teletherapy (MRgRT). The facility is comprised of collinear brachytherapy, MR‐Sim, and linac suites. The MR‐on‐rails translates into the adjacent suites for imaging. An existing bunker and adjacent space were renovated to permit the MR to be moved away from the linac during irradiation. This work describes the design and radiation shielding characteristics of the linac suite within the MRgRT facility. Methods: The original linac vault was designed with a two‐corner maze design to operate at 18 MV. The maze was removed and replaced with three direct‐shielded doors: a sliding console door allow for personnel and gurney access to the bunker; and two double sliding doors through which the MRI translates. The double doors employ an S‐shaped coupling surface to minimize radiation leakage when they are closed. These doors are also RF‐shielded as part of the MR RF cage. The shielding was designed for a workload of 5 × 106 cGy per year; 6 MV beam; maximum dose rate: 1400 MU/min; shielding level 0.05 mSv/year for members of the public and 1 mSv/year for nuclear energy workers. Results: The total thickness of the linac‐MRI sliding doors was 7 inches of lead and 2 inches of stainless steel, producing an equivalent thickness of 8 inches (20.3 cm) of lead. The dose rate at the distal surface of the linac‐MR door is 0.08 mSv/year or 10.9 μSv/hr. The final design includes a copper RF shield and bladder to provide RF shielding. Conclusion: A facility comprising three collinear brachytherapy, MR‐Sim, and linac facility has been designed to meet radiation and radiofrequency shielding limits.

Introduction: An online Magnetic Resonance guided Radiation Therapy (MRgRT) system is under development. The system is comprised of an MRI with the capability of travel between and into HDR brachytherapy and external beam radiation therapy vaults. The system will provide on‐line MRimages immediately prior to radiation therapy. The MRimages will be registered to a planning image and used for image guidance. With the intention of system safety we have performed a failure modes and effects analysis.Methods: A process tree of the facility function was developed. Using the process tree as well as an initial design of the facility as guidelines possible failure modes were identified, for each of these failure modes root causes were identified. For each possible failure the assignment of severity, detectability and occurrence scores was performed. Finally suggestions were developed to reduce the possibility of an event. Results/Discussion: The process tree consists of nine main inputs and each of these main inputs consisted of 5 – 10 sub inputs and tertiary inputs were also defined. The process tree ensures that the overall safety of the system has been considered. Several possible failure modes were identified and were relevant to the design, construction, commissioning and operating phases of the facility. The utility of the analysis can be seen in that it has spawned projects prior to installation and has lead to suggestions in the design of the facility.

Purpose: To evaluate and identify safety system concerns and possible failure modes for a multimodality, linear accelerator(linac) — magnetic resonanceimager(MRI) ‐ brachytherapy,radiotherapy system. Methods: A Delphi process is applied to investigate the safety system concerns and possible failure modes in a linac‐MRI‐brachytherapy system. Well established, for instance in the automotive or airline industries, system, design and process failure modes effect analysis can be applied to the design of a multimodality radiation therapy system. Results: Safety design, systems, processes and culture in radiation therapy is of great importance. When implementing new technologies a review of the necessary staff and patient safety may be necessary. To address patient and staff safety concerns a thoughtful design process must be implemented when an MRI is present or near a linac or a brachytherapy system. Each input into a patient specific process map, such as consent, screening, immobilization, planning, image guidance and machine movement coordination must be analyzed for safety and failure modes concerns. Numerous items have been identified as potential failure modes, these items have been classified into the following nine categories: screening, motion, interlocks, imaging, treatment delivery, dosimetry, plan adaptation, mechanical and miscellaneous. A broad range of failure modes have been identified some of these include and range from non‐compatible in‐room finishing to linac performance degradation under repeated small magnetic field influences to improper patient screening and the possibility of machine collisions. Conclusions: An abridged list of suggestions resulting from our safety systems analysis includes a rigorous staff educational and training program, the evaluation of current quality assurance devices used for mechanical and dosimetric tests, an interlocked computerized system for machine translations and rotations, machine servicing protocols and both fire and patient emergency protocols.

The standard treatment for patients with early stage non-small cell lung cancer is a surgical approach; however inoperable patients may be treated with stereotactic body radiotherapy (SBRT) with an excellent local control rate of 80 to 90%. Although the majority of patients tolerate this treatment reasonably well, some may experience side effects such as chest wall pain and rib fracture. In this study several potential dosimetric and clinical factors related with rib fracture were evaluated and a nomogram estimating the risk of rib fracture based on the most relevant risk factors was created. Our study supported the relationship between dosimetric-clinical factors and rib fractures in patients with lung cancer treated with stereotactic radiotherapy. Based on our findings and supported published data, we have modified our radiotherapy dose in high risk patients for rib fracture.%%%%M.Sc.

The development of image-guidance technologies offer increased precision and accuracy in the placement of radiation therapy dose distributions in the human body. Kilovoltage and megavoltage systems now offer sub-mm accuracy and precision in localizing the treatment and planned isocenters. While this represents a significant improvement in our ability to target lesions, the performance of these systems in terms of their precise and accurate placement of field edges, including individual multi-leaves, is not performing at the same level. The development of dosimeters with small, imageable regions of response will allow these new systems to be evaluated with a much higher level of precision and accuracy. Furthermore, the development of completely automated systems that exploit the robotic nature of the imaging and delivery subsystems should permit the true performance of current IG-IMRT systems to be realized, quantified, and monitored. This includes adjustments to the treatment machines operating parameters, as well as, refinements of the planning system's beam model. Specifically, investigations into novel approaches of employing these detectors, automating their use to maximize system performance, and, the development of novel detectors for compatibility in MR-based systems will be discussed.

Purpose: Palliative Radiotherapy (PRT) plays an important role in the relief of neurologic symptoms in patients with brain metastases; however, little is known about the use of PRT on a population level. The purpose of this study was to describe temporal trends in the use of PRT for brain metastases in Ontario. Methods: The Ontario Cancer Registry was used to gather information on all cancer deaths in Ontario between the years 1984-2004. The proportion of these cases receiving at least one course of PRT for bone metastases within the last two years of life was described over time and by disease site. Results: There were 435,055 cancer deaths in Ontario during the study period. Of these cases, 4.0% received at least one course of PRT for brain mets within the last two years of life. The rate of PRT varied significantly by primary site (e.g. lung8.7%, breast-6.4%, kidney-5.0%, prostate-0.6%) (p

Purpose: Thresholding of 18 F fluorodeoxyglucose (FDG) uptake imaged with positron emission tomography(PET) has been proposed for radiotherapy target delineation. However, there exists no consensus on the threshold best corresponding to disease extent, due to a lack of truth for validation. The purpose of this investigation was to establish criteria under which a fixed FDG‐PET threshold may be useful for target delineation. An image‐based surrogate of target definition ‘truth’ is proposed, derived from a measure of the concordance between multiple expert observers defining the gross tumor volume (GTV) with FDG‐PET and CT. A method was developed that can automatically localize these regions and was used to evaluate the coincident FDG uptake threshold. Methods and Materials: 10 patients with head and neck (H&N) cancers underwent FDG‐PET‐CT imaging using a hybrid scanner (Biograph 16, Siemens Medical Systems). For each patient, 8 observers specializing in H&N cancers delineated the GTV on concurrently‐displayed FDG‐PET‐CT. Regions of high target definition concordance were localized by forming the union of observer target definitions, applying a gradient filter and selecting an intensity threshold equivalent to 6 of 8 observers. At regions of high concordance not attributable exclusively to CT information, the FDG uptake threshold relative to maximum uptake in the GTV was measured. Results: In 4 patients, the mean FDG uptake threshold was 40.9%. In regions of concordance not attributable to CT, the mean threshold in 6 patients was 32.8%. Conclusion: A method was developed for localizing regions of high inter‐observer concordance. These regions served as a surrogate truth for disease extent against which FDG‐PET uptake thresholds were evaluated. For primary H&N cancers, a threshold of approximately 30–40% was able to delineate the tumor at regions of the target boundary where CT information was not definitive. The application of uniform thresholds for H&N target delineation is not recommended.

Gy/2-2.5 weeks), and mid-RT (at 45-50 Gy/4-5 weeks). The time-signal intensity (SI) curve and the plateau SI were derived for each tumor pixel of each DCE MRI. The poorly perfused tumor subregions were quantified as lower 10th percentiles of SI among the tumor pixels (SI10). Outcome endpoints, local tumor control and disease-free survival, were correlated with perfusion parameters of the pre-RT, early RT, and mid-RT MRI. Median follow was 4.6 (range: 0.2-9.0) years. Results: Patterns of DCE of the tumor pixels changed during treatment, and the change influenced therapy outcome: In the pre-RT MRI, 16 patients with initial high perfusion (SI10 $ 2.1) had excellent outcome (100% local control and 81% disease-free survival).In28(34%)ofthe82patientswithinitiallylow perfusion(SI10\2.1) onpre-RTMR, thepatternchangedto highperfusion in early-RT (at 20-25 Gy). These patients had an improved local control rate of 93%, compared to those with persistently low perfusion in pre-RT and early-RT MRIs (72%, p = 0.014). If low perfusion persisted to the mid-RT phase (at 45-50 Gy), local control rate decreased further to 66% (p = 0.008). Patients with persistently low perfusion had a disease-free survival rate of 44%, compared to 81% in those with high pre-RT perfusion, and 72% in those with low pre-RT and subsequently high perfusion (p = 0.01). In patients with initially high perfusion pre-RT, a subsequent decrease in SI did not negatively impact on outcome. Conclusions: Tumor perfusion pattern varies during treatment, and such changes critically influence treatment outcome. Persistently low perfusion from the pre-RT through mid-RT phase indicates high risk of treatment failure, while better outcome is associatedwithhighinitial,orhighsubsequentperfusionpatterns,evenwithpriorstatusofpoorperfusion.Ourlimiteddatasuggest thatDCEMRIisaneffectiveandnon-invasivemeanstoassessthetemporalchangesoftumorperfusionstatusduringthetreatment, which may play an essential role in tumor reoxygenation and radioresponsiveness.

Purpose: To develop an efficient automated method for spine segmentation in cone‐beam CT for online planning of spinal metastases radiotherapy.Method and Materials: A semi‐automated spine segmentation and labeling approach has been developed and validated. The user provides the position and label for an arbitrary vertebra center, and the algorithm automatically identifies the outlines and labels of all visible vertebrae in the field on view. The method is based on adapting 3‐D deformable vertebrae models, placed according to the relative positions derived from a spine atlas. Using the segmentation results, ellipsoid ROIs can be superposed automatically to the vertebrae using co‐registration of the segmented vertebrae with the vertebrae models in the atlas. Ellipsoid ROIs can be used for margin design and avoid excessive expansion due to vertebra processes. Results: The method has been validated on 29 CBCT spine datasets, representing all treatment areas: cervical, thoracic and lumbar. The total number of visible vertebrae was 234. For every dataset, segmentation was started from every visible vertebra. The deviations between the automatically detected position of the vertebra centers and their manual estimations were computed. For about 59% of all vertebrae, the median deviation was less than 5 mm, in approx. 33% of the cases was between 5 and 10 mm, and in approx. 8% of the cases was over 10 mm, which is an indication for a segmentation failure. The running time was a few seconds on a 3.4 GHz PC. Conclusion: Validation study of a one‐click semi‐automated approach to spine segmentation and labeling in cone‐beam CT has been performed. The method has the potential to significantly reduce the time required to identify clinically relevant areas in online planning. A fully automated version of the algorithm is currently under investigation. Conflict of interest: This research is supported in part by Philips Healthcare.

Purpose: To efficiently propagate patient contours to multi‐fraction Cone‐Beam CT volumes. Method and Materials: An auto‐segmentation scheme was developed which uses anchor points to propagate structure delineations from one image to another, accounting for the changing patient anatomy. Four patients involved in 35‐ or 40‐fraction Head‐and‐Neck treatment schemes were selected retrospectively. The 8–12 Organ‐at‐Risk contours drawn in their initial planning CT included spinal cord, brainstem, optic structures, and parotids, to be propagated to the CBCTimage acquired at each fraction. Our method involves three steps: a) Generate a patient‐specific compressed model in terms of salient points from the planning CTimage. The automatic extraction of these uses a developed 3‐dimensional extension of the SIFT algorithm; b) Retrieve this model in every CBCT dataset, via automatic block‐matching maximizing the local correlation between neighborhoods around the interest points; c) Propagate the original contours based on the thin‐plate‐spline warping transformation derived from salient point pairs. Results: For each patient, over 500 salient points were extracted in 5–7 minutes, and subsequently retrieved within 45 seconds in every CBCTimage, into which the 8–12 contours were propagated in 10 seconds. The results show clear improvement, compared to contours obtained by both a mere copy and voxel‐based rigid registration; the typical neck shrinking and spine flexions especially are followed. Our method is more robust and computationally affordable than voxel‐based deformable image registration.Conclusion: The contours can be successfully propagated to all fractions based on significant salient anchor points extracted to summarize one patient's gray‐level information and geometry. The thus highlighted deformations of the critical structures i) can help determine a safe dose map and be reflected in Planning organ at Risk Volume (PRV) margins, ii) may be used to support time‐efficient re‐planning in adaptive radiotherapy.Conflict of Interest: This research is partially supported by Philips Healthcare.

Purpose: To improve the performance of the automatic detection of landmarks, used for 3‐D deformable model initialization, for automatic contouring of the head and neck region in radiation treatment planning. Method and Materials: An automatic landmark detection method has previously been reported by our group. The method uses principal component analysis (PCA) of landmark distributions and 10 nonlinearly registered patient datasets to generate a patient atlas. A controlled random search algorithm (CRS) seeks the optimal transformation between the atlas and the patient dataset to be contoured. The optimal transformation parameters are then used to propagate the atlas landmarks to the patient dataset. In this work, we propose an improved landmark detection method by optimizing the impact of various parameters on the performance of our method and present the results. The parameters involved include varying the number of eigenmodes, increasing the number of landmarks used for nonlinear registration of the atlas datasets, and reducing the constraints imposed by principal component analysis (PCA), by allowing a locally unconstrained search for each landmark. Results: The method has been validated using 10 atlases, each tested on 10 datasets not included in the atlas. Optimization of our automatic landmark identification parameters for the 7 landmarks identified as having the poorest performance, led to a mean rms‐distance to ground truth of 11.2 ± 1.5 mm (range: 7.7 – 18.2 mm), an improvement of 5.9 mm (30%) compared with previously reported results. Conclusion: We optimized the impact of 14 parameters on the performance of automatic landmark identification. As a result, we have reduced the mean rms‐distance to ground truth by 5.9 mm. Improved model initialization through optimized landmark detection, will result in more robust model adaptation and thereby improved automated contouring of the organs at risk. Conflict of Interest: This work is supported in part by Philips Healthcare.

Purpose: The clinical implementation of conformal radiotherapyIMRT has substantially increased the need for automated contouring tools. Implementation of image‐guidedadaptive RT will further increase the need. The objective of this work is to develop and assess a fully automated method for delineating risk structures on the basis of statistical models and anatomical atlases. Method and Materials: Anatomical variability is modeled from a set of training images from multiple patients, where structures of interest were manually contoured by clinical experts, and salient anatomical landmarks were defined, e.g. the mastoid process. We use landmark correspondences and thin‐plate spline (TPS) approximation to map all training data into a common geometrical reference frame and generate a mean volumetric image and mean 3D surfaces; spatial variation and dependency of the landmarks is encoded in a statistical model and their local appearance is derived from the mean image. This model is then used to automatically find the positions of the landmarks in a new patient image. TPS is used to map the contours from the mean volume onto the new volume. Finally, the mapped contours are refined using a model‐based segmentation technique. Results: We evaluated our method for three anatomical regions: head‐neck, pelvis and thorax, on multiple clinical datasets accompanied with manually created contours. Quantitative results where collected for spatial accuracy, both for detection of the landmarks as well as for the final contours. For initialization the distance to ground‐truth was 8.9 ± 4.8 mm, contour distance was 3.5 ± 4.5 mm across all anatomies. Computation time was below 50 seconds in all cases. Conclusion: A fully automated segmentation method was developed and quantitatively evaluated for risk structures from CTimages in the head‐neck, pelvis and thorax. Future work will increase consistency and efficiency for structures with poor image contrast.