- Email: [email protected]
Technique for Targeting Arteriovenous Malformations Using Frameless Image-Guided Robotic Radiosurgery
Int. J. Radiation Oncology Biol. Phys., Vol. 79, No. 4, pp. 1232–1240, 2011 Copyright Ó 2011 Elsevier Inc. Printed in the USA. All rights reserved 0360-3016/$–see front matter
doi:10.1016/j.ijrobp.2010.05.015
PHYSICS CONTRIBUTION
TECHNIQUE FOR TARGETING ARTERIOVENOUS MALFORMATIONS USING FRAMELESS IMAGE-GUIDED ROBOTIC RADIOSURGERY DIMITRE HRISTOV, PH.D.,* LINA LIU, M.SC.,* JOHN R. ADLER, M.D.,*y IRIS C. GIBBS, M.D.,* TERI MOORE,{ MARILY SARMIENTO,{ STEVE D. CHANG, M.D.,y ROBERT DODD, M.D.,yz MICHAEL MARKS, M.D.,yz AND HUY M. DO, M.D.yz *Radiation Oncology, yNeurosurgery, and zDiagnostic Radiology, Stanford University School of Medicine, Stanford, California, and { Siemens Medical Solutions, Malvern, Pennsylvania Purpose: To integrate three-dimensional (3D) digital rotation angiography (DRA) and two-dimensional (2D) digital subtraction angiography (DSA) imaging into a targeting methodology enabling comprehensive image-guided robotic radiosurgery of arteriovenous malformations (AVMs). Methods and Materials: DRA geometric integrity was evaluated by imaging a phantom with embedded markers. Dedicated DSA acquisition modes with preset C-arm positions were configured. The geometric reproducibility of the presets was determined, and its impact on localization accuracy was evaluated. An imaging protocol composed of anterior-posterior and lateral DSA series in combination with a DRA run without couch displacement between acquisitions was introduced. Software was developed for registration of DSA and DRA (2D-3D) images to correct for: (a) small misalignments of the C-arm with respect to the estimated geometry of the set positions and (b) potential patient motion between image series. Within the software, correlated navigation of registered DRA and DSA images was incorporated to localize AVMs within a 3D image coordinate space. Subsequent treatment planning and delivery followed a standard image-guided robotic radiosurgery process. Results: DRA spatial distortions were typically smaller than 0.3 mm throughout a 145-mm 145-mm 145-mm volume. With 2D-3D image registration, localization uncertainties resulting from the achievable reproducibility of the C-arm set positions could be reduced to about 0.2 mm. Overall system-related localization uncertainty within the DRA coordinate space was 0.4 mm. Image-guided frameless robotic radiosurgical treatments with this technique were initiated. Conclusions: The integration of DRA and DSA into the process of nidus localization increases the confidence with which radiosurgical ablation of AVMs can be performed when using only an image-guided technique. Such an approach can increase patient comfort, decrease time pressure on clinical and technical staff, and possibly reduce the number of cerebral angiograms needed for a particular patient. Ó 2011 Elsevier Inc. AVM, Radiosurgery, Image guidance, Cyberknife, Angiography, Registration.
Two technologies allow a different, frameless approach to AVM radiosurgery. C-arm CT (CACT, also termed digital rotation angiography [DRA]) generates cone beam CT anatomical and vascular data in the DSA imaging setting, thus facilitating the correlation of 3D and 2D vasculature datasets for treatment planning. A robotic delivery system (Cyberknife; Accuray Inc, Sunnyvale, CA) uses frequent stereoscopic X-ray guidance to continuously localize the target and, in response to patient movements, redirect the beam during delivery. Based on these technologies, a frameless AVM radiosurgical treatment process can be envisaged that consists of the following steps: (1) a localization neuroangiography
INTRODUCTION Stereotactic radiosurgery is an important component of a multimodal arteriovenous malformation (AVM) treatment that historically uses a stereotactic frame rigidly fixed to a patient’s head. For treatment planning, imaging with a mode-specific localizer attached to the frame enables fusion of computed tomography (CT)/magnetic resonance (MR) and digital subtraction angiography (DSA) datasets, the last being essential for defining the nidus of some AVMs (1,2). Throughout radiosurgical treatment, attachment of the frame to the treatment device assures accurate localization and robust immobilization of the target. Reprint requests to: Dimitre Hristov, Ph.D., 875 Blake Wilbur Dr. Stanford, CA 94305-5847. Tel: (650) 498-7898; Fax: (650) 4984015; E-mail: [email protected] This work was presented in part at the 51st Annual Meeting of American Society for Therapeutic Radiology and Oncology, Nov 1–5, 2009, Chicago, IL.
Conflict of interest: Teri Moore and Marily Sarmiento are employees of Siemens Medical Solutions. John R. Adler is a shareholder of Accuray, Inc. Received Jan 7, 2010, and in revised form May 10, 2010. Accepted for publication May 14, 2010. 1232
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
procedure to acquire DRA and DSA datasets; (2) DSA-DRA (2D-3D) registration for spatial correlation of DSA and DRA image features; (3) AVM localization within the DRA image space by spatially linked navigation within the DRA and DSA datasets; (4) acquisition of a planning CT followed by DRA-CT registration and transfer of the DRA-localized target onto the CT image space; and (5) standard treatment planning and subsequent image-guided targeting utilizing Cyberknife. Colombo et al. (3) recently reported early results of Cyberknife radiosurgery using some of the methods proposed above. Significantly, frameless AVM localization based solely on DRA images and DRA/CT/MR registration was used (3). However, for AVM nidus definition, DRA by itself does not provide all the anatomic and physiologic information that comes with ‘‘time-stamped’’ DSAs. Therefore, for a comprehensive approach to frameless nidus localization, the incorporation of DSA is essential in many cases (1, 2). Thus, the purpose of the present project was to develop a process for frameless AVM localization consisting of steps 1 to 3, outlined above, and to evaluate the associated localization uncertainty.
METHODS AND MATERIALS C-arm DRA imaging A flat-panel C-arm imaging system (Axiom Artis; Siemens Medical Solutions, Inc, Malvern, PA) supporting various DRA modes was used. A DRA scan involves continuous C-arm rotation on a circular trajectory around a subject with concurrent acquisition of Xray projection images. The acquisition geometry for each DRA mode is calibrated first by scanning a PDS-2 calibration phantom (Siemens AG, Forchheim, Germany) with 108 spherical fiducial markers (Fig. 1A) embedded in a cylinder with a diameter of 145 mm. The axial extent of the fiducial pattern is about 145 mm. The known 3D marker positions in the coordinate system of the phantom and the 2D positions of the marker projections (Fig. 1B) in the coordinate systems of the individual CACT X-ray projections are used to derive 3 4 projection matrices which define a unique correspondence between a 3D point and its 2D projection (Fig. 1C) (4). These matrices are used in the DRA reconstruction for backprojection of the filtered projection images.
1233
The available DRA modes offer different compromises between acquisition time, image resolution, and contrast. For frameless AVM localization, we selected the ‘‘5s-1k DR 480 1.5 00.36’’ mode acquiring 126 projection images over 188 in 5 seconds for the reconstruction of pre- and postcontrast DRA volumes encompassing 256 256 222 voxels at 0.81-mm isotropic voxel size. While providing excellent visualization of contrast-filled cerebral vessels and high-contrast bony anatomy (Fig. A. 1, Supplementary Material), this mode also minimizes the probability of patient motion during image acquisition and maximizes the DRA reconstruction field-of-view to facilitate subsequent DRA-CT registration based on bony anatomy.
Spatial integrity of DRA and CT images For localization, DRA datasets need to accurately depict the spatial relationships between anatomical features. However, cone beam images reconstructed from a circular trajectory exhibit artifacts due to the lack of sufficient information for accurate 3D reconstruction ((5) and references therein). Furthermore, other potential geometric distortions may arise as a result of variances in the calibration and mechanical stability of the C-arm. The DRA spatial integrity was evaluated as follows. A DRA scan of the calibration phantom (Fig. 1A) was acquired and reconstructed. CT phantom images were also acquired according to a local radiosurgical simulation protocol: axial scanning; slice thickness, 1.25 mm; pixel size, 0.55 mm. The fiducial centers (Fig. 1A) within the DRA and the CT volumes were determined by intensity-based center-of-mass calculation within a volume-of-interest (VOI) encompassing each fiducial marker. The VOIs around the fiducials were segmented by a global, user-defined threshold. The lengths of segments connecting the centers of all possible fiducial pairs were calculated (Fig. 2). The mean and the standard deviation of the differences between the image-based segment lengths and the segment lengths derived from the phantom specifications served as metrics for the DRA and CT spatial integrity values.
Localization imaging protocol: description and reproducibility In order to facilitate the spatial correlation of DSA and DRA datasets, we introduced an imaging protocol consisting of DSA anteriorposterior (AP), DSA lateral (LAT), and DRA series with fixed couch position for all series. The parameters for these acquisitions were entered as preset programs in the imaging system to assure consistency and reproducibility of the acquisition geometries. The AP and LAT C-arm positions were selected by matching AP/LAT calibration
Fig. 1. (A) 3D rendering of the C-arm CT calibration phantom. (B) A projection image of the calibration phantom demonstrating the embedded fiducial markers. (C) 3D-2D point correspondence described by a projection matrix P.
1234
I. J. Radiation Oncology d Biology d Physics
Volume 79, Number 4, 2011
At the core of Angio-Localize is the procedure for registering DSA and DRA images detailed below.
DSA-DRA registration
Fig. 2. 3D rendering of the C-arm CT calibration phantom demonstrating the calculated centers of the fiducial markers. Sample segments connecting the centers of fiducial pairs are shown. phantom images to corresponding projections from the DRA series. Thus, provided that accuracy of the match is sufficient and consistently reproducible, the 3 4 matrices associated with the DRA projections could be used to link the DRA image space to the AP/LAT DSA series. We evaluated the geometric reproducibility of the DSA presets in two imaging sessions separated by 11 days. During each session, several AP and LAT views (Fig. 1B) of the calibration phantom were acquired. Prior to each series, the C-arm was parked and then placed in an arbitrary position and then instructed to move to the preset imaging position. The centers of the fiducial markers within the images (Fig. 1B) were determined by intensity-based centerof-mass calculation within a region-of-interest (ROI) encompassing each fiducial. The ROIs around the fiducials were segmented by a global, user-defined threshold. Overlapping fiducials clustered in the periphery of the radiographic projections (Fig. 1B) were excluded from the analysis, since their centers could not be reliably extracted by center-of-mass calculations. For the individual fiducial markers, the absolute deviations in position captured across a series of images were calculated. For consistency with the metrics used for DSA-DRA registration, the maximum values of the absolute deviations and population means of the absolute deviations were used as image-based metrics for geometry reproducibility.
Angio-Localize software We developed software (Angio-Localize) to enable integration, visualization, and exploration of various 3D and 2D datasets for frameless AVM localization. The application is based on a freely available medical visualization platform (6) and open source packages. Thus, the application can be deployed and enhanced by new users (http://www.stanford.edu/dhristov). Within the software, a typical workflow for frameless AVM localization is supported by five successively executed task-pad mini-applications (Table 1).
Alignment of the DSA (2D) and DRA (3D) datasets is a 2D-3D registration problem which requires the determination of a 3 4 ! projection matrix P such that for any point X within the DRA 3D ! coordinate space and its 2D projection x depicted in the DSA image ! (Fig. 1C), the equality lx ! x ¼ P X applies, where the scaling coef! ficient lx depends on X . Intensity-based (7, 8), feature-based (8–10), and hybrid (11, 12) techniques have been investigated for 2D–3D vascular registration. We have opted for a semiautomatic feature-based method based on segmented vascular structures (10). Figure 3 illustrates the 2D-3D registration process. Adjusting the viewing parameters used for DRA volume rendering, users interactively specify threshold levels for binary segmentation of 3D vasculature structures within a VOI (Fig. 3A). The centerline of the segmented vasculature is automatically extracted (13) and projected onto the DSA sequence (Fig. 3B). After reviewing the overlay of the projection with the DSA images, users select a particular DSA frame for 2D-3D registration and proceed by specifying threshold levels for segmenting DSA 2D vascular structures (Fig. 3C) within interactively selected ROI. The centerline of the segmented 2D vasculature is then automatically extracted (13) and converted to a distance map image (14) in which pixel values represent distance to the nearest point of the vascular centerline. Before proceeding with the automatic registration of the DRA-segmented vessel tree and the DSA-derived distance map image, users can exclude certain regions of the vascular tree from the registration process by defining ROIs around the projections of these regions on the distance map (Fig. 3D). The automatic 2D-3D registration is performed by solving a minimization problem with respect to the matrix parameters. The cost function is the mean distance between the projection of the DRAsegmented vessel tree centerline and the DSA-segmented vessel tree centerline. The cost is calculated by projecting the points from the 3D centerline onto the 2D distance map and accumulating the pixel values at the locations of the projected points. The matrix, P, can be decomposed as P = KC where K (intrinsic parameters) describes the projection geometry from the X-ray source to the flat-panel detector and C (extrinsic parameters) is a 4 4 affine transformation given as 2 3 tx 6 R ty 7 7; C ¼ ½R j T; T ¼ tx ; ty ; tz 0 : C¼6 4 tz 5 0 0 0 1 The extrinsic parameters describe the rotation, R, and translation, T, of the X-ray source in the DRA space. With square pixels of known dimensions, K can be written as 3 2 SID 0 u0 0 SID v0 0 5 K¼4 0 0 0 1 0 where SID is the source-to-imager distance, and u0 and n0 specify the image coordinates (in mm) of the piercing point of the normal from that of the X-ray source and the imaging plane. As a result of the 2D-3D registration, all 9 parameters (3 rotations, 3 translations, 3 intrinsic) need to be determined. However, small SID variations affect point projection coordinates only slightly. Furthermore, misestimation of u0 and n0 values can be somewhat compensated by adjustments in the X-ray source position. Thus, in our
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
1235
Table 1. Angio-Localize task pad mini-applications and their functionality Mini-application
Functionality
1. Patient selection 2. Image selection
Users create, select and manage records in a patient database. For a selected patient, users import 3D (CT, DRA, MR) and 2D (DSA) images into the database via DICOM interface. Subtraction of post- and pre-contrast datasets is performed. For a selected patient, users perform direct 3D-3D (DRA-CT, CT-MR) and 2D-3D (DSA-DRA) registration. Users review and approve registrations. Some datasets are further registered indirectly by concatenation of user-approved registrations. For a selected patient, users concurrently visualize and navigate registered 3D (CT, DRA, MR) and 2D (DSA) datasets. Users contour an AVM on a 3D dataset with visual feedback from the AVM display on other registered images. Users export AVM contours for treatment planning.
3. Registration 4. Localization 5. Export
implementation, the projection matrix is approximated as P = K0C0 [DR j DT]. In this expression K0 and C0 are fixed and estimated with information from the DICOM headers of the images, and [DR j DT] is a 4 x 4 affine matrix to account for small C-arm misalignments
(rotations, DR; and translations, DT) from the set positions and/or patient movements between image acquisitions. The objective function is thus minimized by a simplex method (15) with respect to the six parameters that determine [DR j DT].
Fig. 3. Steps in the DSA-DRA registration process. (A) Interactive segmentation of the DRA-defined 3D vasculature within a user-defined VOI. (B) Projection of points (visualized as markers) representing the centerline of the segmented 3D vasculature onto a DSA frame from the DSA series. (C) Interactive threshold segmentation of the DSA and overlay with the projected markers. A ROI (black outline) is defined for the subsequent processing steps. (D) Overlay of the projected markers with a distance map image derived from the segmented DSA. Regions (in black) are defined that contain marker outliers to be excluded from the automatic registration that minimizes the average image intensity across all markers. (E) Overlay of the projected markers with the distance transform image for visual inspection of the automatic registration results. (F) Overlay of the projected markers with user-selected DSA frame for visual inspection of the automatic registration results and interactive adjustment.
1236
I. J. Radiation Oncology d Biology d Physics
Volume 79, Number 4, 2011
Fig. 4. 2D-3D registration of DRA-segmented phantom fiducials to a radiograph. (A) Overlay of the projected fiducials and the radiograph that demonstrates initial misregistration. (B) Overlay of the projected fiducials with the radiographderived distance map used for the automatic 2D-3D registration. (C) As in panel B, after the registration. Fiducials excluded from the registration process are encompassed by user-defined ROIs (in black). (D) Overlay of the projected fiducials and the radiograph for visual confirmation of the registration results.
After the registration, P = K0C0[DR j DT] is used to project the 3D vessel tree onto the 2D distance map (Fig. 3E) and onto the DSA frames (Fig. 3F) to allow users to evaluate the registration quality. The cost and the number of centerline points used for its evaluation are reported. Furthermore, additional adjustment of the registration is possible by interactive manipulation of the parameters that determine [DR j DT] and visual evaluation of the alignment between the projection of the 3D vessel centerline and the DSA vasculature (Fig. 3F). We evaluated the inherent accuracy of the 2D-3D registration by registering a DRA scan of the calibration phantom to the AP and LAT views used for the evaluation of the imaging geometry reproducibility in session 2 (as described in Localization imaging protocol: description and reproducibility). An example is shown in Fig. 4. The centers of the calibration phantom fiducials (segmented as described in Spatial integrity of DRA and CT images; see also Fig. 1A and 2) were used as ‘‘segmented vasculature.’’ The residual cost at the end of the automatic registration was the 2D-3D registration metrics. This metric captures the quality of alignment between the fiducial projections and a distance map image. However, it does not consider the additional intrinsic registration uncertainty result-
ing from the discrete distance assignment when radiographs are converted into a distance map. Regardless of other factors, the discrete distance assignment places a lower limit of about 0.5 pixels on the achievable registration accuracy at the detector plane.
DSA-CT/MR registration and localization Within the Angio-Localize software, 3D-3D registration is used to link CT and MR volumes to pre-and postcontrast DRA datasets. The resulting transformations are concatenated with the DRA-toDSA registration matrices to generate perspective transformations that link CT and MR volumes to the DSAs. Once approved by a physician, these 3D-3D and 2D-3D transformations are used within the Localization mini-task for spatially correlated simultaneous navigation of 2D and 3D datasets (Fig. 5). In this process, as one follows the cerebral vessels on fused postcontrast DRA and contrastenhanced CT, the current 3D cursor location is projected on the registered LAT and AP DSA sequences. Similarly, AVM contours drawn on the 3D datasets are concurrently projected onto the DSA images for visual feedback. After the AVM segmentation, a DRA with an annotated AVM target is exported in DICOM format for planning.
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
1237
Fig. 5. AVM localization and contouring by correlated navigation of fused CT and postcontrast DRA datasets (top) and LAT and AP DSA dynamic sequences (bottom) within the Localization mini-task panel of Angio-Localize.
RESULTS Spatial integrity of DRA and CT images Table 2 summarizes the results from the evaluation of the DRA and CT spatial integrity. The systematic differences between image-derived and manufacturer-specified interfiducial distances were smaller than 0.05 mm (Table 2). Over the sample field of view defined by the markers, for the majority of the segments, the image-based interfiducial distances matched the manufacturer’s specifications within 0.5 mm (Table 2). The largest discrepancies were 0.6 mm for the DRA and 0.9 mm for the CT. Localization imaging protocol: geometry reproducibility Table 3 summarizes the results from the evaluation of the geometry reproducibility of the preset localization imaging Table 2. Differences between lengths of corresponding segments extracted from different datasets Parameter Segment length differences* No. of segments
DRA vs. phantom CT vs. phantom specifications specifications
DRA vs. CT
0.0 0.3
0.0 0.5
0.1 0.6
5,565
5,565
5,356
* Data are presented as means 2 standard deviation (mm).
modes. The mean deviations of the marker positions varied between 0.4 mm and 1.5 mm. Deviations as large as 3.2 mm were observed. These may have resulted either from intrinsic limits on the system’s mechanical stability or from operator variability in driving the system toward the set positions, since premature termination of the system movement is possible even if a set position is not reached. We did not investigate this further. Instead, we opted for the development of 2D-3D registration option in the AngioLocalize software to correct for C-arm misalignments from Table 3. Deviations observed in the image-based investigations of the geometric reproducibility of the preset C-arm modes* Session Mean absolute Maximum No. of images (preset) deviation (mm) deviation (mm) (no. of markers/image) 1 (AP) 1 (LAT) 2 (AP) 2 (LAT)
0.4 0.4 1.5 0.8
1.2 1.3 3.2 1.7
5 (51) 5 (44) 6 (48) 6 (44)
* Data show summary of the deviations observed in the imagebased investigations of the geometric reproducibility of the preset C-arm modes. The deviations are specified in the flat-panel detector plane. At the C-arm isocenter, these deviations are demagnified by a factor of 0.625.
1238
I. J. Radiation Oncology d Biology d Physics
Fig. 6. Lateral DSA image from left vertebral arteriogram reveals a tiny residual AVM nidus fed be the parietal occipital branch with early cortical draining vein to the superior sagittal sinus (arrow).
the set positions and/or patient movements between image acquisitions. DSA-DRA registration accuracy After the registration of the calibration phantom DRA scan to the six LAT radiographs in session 2, the residual cost function value calculated over 68 markers varied between 0.02 mm and 0.03 mm in the flat-panel detector plane. For the six AP radiographs, the residual cost function calculated over 58 markers varied between 0.02 mm and 0.04 mm in the detector plane. Thus, the discrete distance assignment in the distance map image is the dominant source of intrinsic registration uncertainty, about 0.5 pixels in the detector plane. With a pixel size of 0.308 mm and conservative radiographic magnification factor of 0.75 (typically 0.625 at C-arm isocenter), this registration uncertainty translates to inherent localization error of O3 0.5 0.308 0.75 mm = 0.2 mm or less within the DRA space. Exemplary clinical case The frameless technique was introduced for the treatment of a 23-year-old woman who required radiosurgery for a tiny postoperative residual AVM. This patient experienced one transient episode of left-sided weakness and numbness 3 years after undergoing surgical resection of a 3-cm-diameter, right posterior temporal parietal AVM, fed by the right posterior cerebral artery (PCA) and draining superficially into the superior sagittal sinus. Preradiosurgical angiography consisting of DRA and DSA acquisitions was performed with the patient under conscious sedation following the localization protocol described in Materials and Methods. During angiog-
Volume 79, Number 4, 2011
raphy, the patient’s head was immobilized by a Velcro strap. Both internal carotid and the dominant left vertebral arteries were selectively catheterized, and biplane digital subtraction angiographies were performed. The right posterior communicating artery was large with robust flow through the right posterior cerebral territory. A small, less than 1 cm, residual AVM nidus was identified, fed by a small parietal-occipital branch of the right PCA and with early drainage by way of a cortical vein and the posterior superior sagittal sinus, (Fig. 6). Subsequently, while immobilized in a head rest and custom face mask, the patient was imaged by using (1.25-mm) thin slice, contrast-enhanced axial CT using a pixel size of 0.55 mm. After approving the DSA-DRA registration (Fig. 7A), the treating neurosurgeon delineated the AVM nidus on the postcontrast DRA image set by concurrent linked navigation of the DRA and DSA images (Fig. 5). The postcontrast DRA with annotated AVM location was then exported to the treatment planning system (Multiplan version 2.2; Accuray Inc.). Treatment planning and delivery followed a standard process. Briefly, the DRA and the CT volume were automatically registered and fused to establish the position of the AVM within the planning CT (Fig. 7B). The commercial registration algorithm is based on normalized mutual information and a hierarchical optimization strategy, broadly similar to the technique described in the report by Studholme et al. (16, 17). A treatment plan was then generated to deliver 20 Gy in a single fraction to the nidus (Fig. 7B). Reference pairs of digitally reconstructed radiographs (DRRs) capturing different potential poses of the patient skull were generated by Cyberknife’s system. Prior to delivery, the patient was remobilized in the same custom accessories used with the planning CT. Then the target was brought to the treatment position by acquiring stereoscopic X-ray images, registering these to reference DRRs and adjusting the treatment table with the transformation parameters derived from registration. During the 90-min radiosurgery, the delivery system acquired an image pair approximately every 2 min. These images were concurrently correlated to the reference DRRs to redirect the radiation beam and track the target. DISCUSSION For intracranial targets that can be accurately localized within the coordinate space of high-resolution CT datasets, image-guided radiosurgery is a convenient alternative to frame-based approaches. Indeed, when assessed with the hidden target, Winston-Lutz test (18) in anthropomorphic phantoms, modern image-guided radiosurgery systems demonstrate targeting accuracy from 0.4 mm to 1.1 mm depending on the system and resolution of the CT scan used for target localization (19–21). Comparable targeting accuracy in the range of 0.3 mm to 1.3 mm has been reported with frame-based approaches (22, 23). Furthermore, frequent intrafractional imaging combined with skull-based tracking can reduce random targeting errors from patient motion
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
1239
Fig. 7. (A) Overlay of the projected segmented 3D vasculature with a frontal DSA and a lateral DSA for review and approval of the DSA-DRA registration. (B) Fused planning CT and postcontrast DRA images overlaid with the planned isodose distribution. The prescription dose is 20 Gy to the region encompassed by 80% (isodose line in green) of the maximum dose (25 Gy).
down to 1 mm (24), which is similar to the motion-induced error with frame-based radiosurgery (21). However, a comprehensive frameless radiosurgery technique for AVMs requires the integration of DSA datasets in the localization process (1, 2), which is the significance of the current work. We opted for a geometric, featurebased semiautomatic DSA-DRA registration, since the cost function indicates the potential localization error resulting from residual 2D-3D misalignment. While such a cost function was previously used (10), our implementation differs in two aspects. First, a faster optimization scheme is applied because the DSA and the DRA datasets are already approximately aligned as a result of the imaging protocol. Second, tools are provided for users to interactively remove outliers that can bias the registration. Such adjustments significantly improve the robustness of the registration process,
which is still contained within 5 to 10 min for an experienced user. The system-related localization error within the DRA space resulting from geometric distortions (Table 2) and automatic 2D-3D registration uncertainty (0.2 mm) is conservatively about 0.4 mm, thus comparable to a reported stereotactic targeting error of 0.4 0.1 mm based on biplanar radiographic imaging (22). Clinically, however, the localization error is determined by the accuracy of the physician-adjusted and -reviewed DSA-DRA and DRACT registrations. In our experience, during interactive 2D-3D registration, shifts as small as 0.2 mm in directions parallel to the imaging plane are readily identifiable through the inspection of the overlay of the projected vessel centerline and the DSA frames (Figs. 3F, 7A, and A. 2, Supplementary Material). On the other hand, even though the
1240
I. J. Radiation Oncology d Biology d Physics
DRA-CT registration is facilitated by the similar appearance of bony anatomy and the presence of some common vascular landmarks in the DRA and the contrast-enhanced CT (Fig. A.1), the local error is typically about 1 mm as evaluated from homologous landmarks in the images (Fig. A.1, Supplementary Material). Thus, the DRA-CT registration dominates the system-related uncertainty in the frameless AVM localization process. However, new angiography systems are capable of large field-of-view DRA imaging, and thus, DRA datasets can consistently capture the entire cranial anatomy. Thus, the DRA images may potentially be used on their own for the generation of DRRs as well as dose-calculation after appropriate electron density calibration.
Volume 79, Number 4, 2011
CONCLUSIONS In summary, the availability of DRA and image-guided delivery technology enables a comprehensive frameless radiosurgery process for AVMs. With this process, patient comfort is increased due to the elimination of pain or potential scar formation from placement of a stereotactic frame. Time pressure on the staff is reduced, since imaging, planning, and treatment sessions do not have to be completed in 1 day. The number of angiograms for a particular patient can also be potentially reduced by including DRA acquisition as part of the diagnostic imaging process, thereby eliminating the need for a subsequent imaging procedure used specifically for treatment planning.
REFERENCES 1. Zhang XQ, Shirato H, Aoyama H, et al. Clinical significance of 3D reconstruction of arteriovenous malformation using digital subtraction angiography and its modification with CT information in stereotactic radiosurgery. Int J Radiat Oncol Biol Phys 2003;57:1392–1399. 2. Hamm KD, Klisch J, Surber G, et al. Special aspects of diagnostic imaging for radiosurgery of arteriovenous malformations. Neurosurgery 2008;62:A44–52. 3. Colombo F, Cavedon C, Casentini L, et al. Early results of CyberKnife radiosurgery for arteriovenous malformations. J Neurosurg 2009 Oct;111(4):807–819. 4. Navab N, Bani-Hashemi A, Nadar MS, et al. 3D reconstruction from projection matrices in a C-arm based 3D-angiography system. Medical Image Computing and Computer-Assisted Intervention -. Miccai’98 1998;1496:119–129. 5. Bartolac S, Clackdoyle R, Noo F, et al. A local shift-variant Fourier model and experimental validation of circular conebeam computed tomography artifacts. Med Phys 2009;36: 500–512. 6. MeVisLab. Adevelopment environment for medical image processing and visualization, MeVis. Available at: http://www. MeVisLab.de. 2006. Accessed July 13, 2010. 7. Bullitt E, Liu A, Aylward SR, et al. Registration of 3D cerebral vessels with 2D digital angiograms: Clinical evaluation. Acad Radiol 1999;6:539–546. 8. McLaughlin RA, Hipwell J, Hawkes DJ, et al. A comparison of a similarity-based and a feature-based 2-D-3-D registration method for neurointerventional use. IEEE Trans Med Imaging 2005;24:1058–1066. 9. Groher M, Zikic D, Navab N. Deformable 2D-3D registration of vascular structures in a one view scenario. IEEE Trans Med Imaging 2009;28:847–860. 10. Florin C, Williams J, Khamene A, et al. Registration of 3D angiographic and X-ray images using sequential Monte Carlo sampling. Computer Vision for Biomedical Image Applications. Proceedings 2005;3765:427–436. 11. Jomier J, Bullitt E, Van Horn M, et al. 3D/2D model-to-image registration applied to TIPS surgery. In Larsen R, Nielsen M, Sporring Jon (editors). Medical Image Computing and Computer-Assisted Intervention - MICCAI 2006: 9th International Conference, Copenhagen, Denmark, October 2006 Proceedings. 2006;4191:662–669.
12. Vermandel M, Betrouni N, Pasquier D, et al. A 2D/3D matching based on a hybrid approach: improvement to the imaging flow for AVM radiosurgery. Conf Proc IEEE Eng Med Biol Soc 2005;3:3071–3073. 13. Selle D, Peitgen HO. Analysis of the morphology and structure of vessel systems using skeletonization. In: Chen C, Clough A, editors. Proc SPIE 2001;4321;. p. 271–281. 14. Fitzgibbon AW. Robust registration of 2D and 3D point sets. Image Vis Comput 2003;21:1145–1153. 15. Nelder JA, Mead R. A simplex-method for function minimization. Comp J 1965;7:308–313. 16. Studholme C, Hill DL, Hawkes DJ. Automated threedimensional registration of magnetic resonance and positron emission tomography brain images by multiresolution optimization of voxel similarity measures. Med Phys 1997;24:25–35. 17. Maurer CR Jr., West JB. Medical image registration using mutual information. In: Heilbrun MP, editor. Cyberknife radiosurgery: Practical guide 2. Sunnyvale, CA: The Cyberknife Society; 2006. p. 23–34. 18. Lutz W, Winston KR, Maleki N. A system for stereotactic radiosurgery with a linear accelerator. Int J Radiat Oncol Biol Phys 1988;14:373–381. 19. Antypas C, Pantelis E. Performance evaluation of a CyberKnife G4 image-guided robotic stereotactic radiosurgery system. Phys Med Biol 2008;53:4697–4718. 20. Lamba M, Breneman JC, Warnick RE. Evaluation of imageguided positioning for frameless intracranial radiosurgery. Int J Radiat Oncol Biol Phys 2009;74:913–919. 21. Solberg TD, Medin PM, Mullins J, et al. Quality assurance of immobilization and target localization systems for frameless stereotactic cranial and extracranial hypofractionated radiotherapy. Int J Radiat Oncol Biol Phys 2008;71:S131–S135. 22. Yeung D, Palta J, Fontanesi J, et al. Systematic analysis of errors in target localization and treatment delivery in stereotactic radiosurgery (SRS). Int J Radiat Oncol Biol Phys 1994;28:493– 498. 23. Rahimian J, Chen JC, Rao AA, et al. Geometrical accuracy of the Novalis stereotactic radiosurgery system for trigeminal neuralgia. J Neurosurg 2004;101(Suppl 3):351–355. 24. Murphy MJ. Intrafraction geometric uncertainties in frameless image-guided radiosurgery. Int J Radiat Oncol Biol Phys 2009;73:1364–1368.
doi:10.1016/j.ijrobp.2010.05.015
PHYSICS CONTRIBUTION
TECHNIQUE FOR TARGETING ARTERIOVENOUS MALFORMATIONS USING FRAMELESS IMAGE-GUIDED ROBOTIC RADIOSURGERY DIMITRE HRISTOV, PH.D.,* LINA LIU, M.SC.,* JOHN R. ADLER, M.D.,*y IRIS C. GIBBS, M.D.,* TERI MOORE,{ MARILY SARMIENTO,{ STEVE D. CHANG, M.D.,y ROBERT DODD, M.D.,yz MICHAEL MARKS, M.D.,yz AND HUY M. DO, M.D.yz *Radiation Oncology, yNeurosurgery, and zDiagnostic Radiology, Stanford University School of Medicine, Stanford, California, and { Siemens Medical Solutions, Malvern, Pennsylvania Purpose: To integrate three-dimensional (3D) digital rotation angiography (DRA) and two-dimensional (2D) digital subtraction angiography (DSA) imaging into a targeting methodology enabling comprehensive image-guided robotic radiosurgery of arteriovenous malformations (AVMs). Methods and Materials: DRA geometric integrity was evaluated by imaging a phantom with embedded markers. Dedicated DSA acquisition modes with preset C-arm positions were configured. The geometric reproducibility of the presets was determined, and its impact on localization accuracy was evaluated. An imaging protocol composed of anterior-posterior and lateral DSA series in combination with a DRA run without couch displacement between acquisitions was introduced. Software was developed for registration of DSA and DRA (2D-3D) images to correct for: (a) small misalignments of the C-arm with respect to the estimated geometry of the set positions and (b) potential patient motion between image series. Within the software, correlated navigation of registered DRA and DSA images was incorporated to localize AVMs within a 3D image coordinate space. Subsequent treatment planning and delivery followed a standard image-guided robotic radiosurgery process. Results: DRA spatial distortions were typically smaller than 0.3 mm throughout a 145-mm 145-mm 145-mm volume. With 2D-3D image registration, localization uncertainties resulting from the achievable reproducibility of the C-arm set positions could be reduced to about 0.2 mm. Overall system-related localization uncertainty within the DRA coordinate space was 0.4 mm. Image-guided frameless robotic radiosurgical treatments with this technique were initiated. Conclusions: The integration of DRA and DSA into the process of nidus localization increases the confidence with which radiosurgical ablation of AVMs can be performed when using only an image-guided technique. Such an approach can increase patient comfort, decrease time pressure on clinical and technical staff, and possibly reduce the number of cerebral angiograms needed for a particular patient. Ó 2011 Elsevier Inc. AVM, Radiosurgery, Image guidance, Cyberknife, Angiography, Registration.
Two technologies allow a different, frameless approach to AVM radiosurgery. C-arm CT (CACT, also termed digital rotation angiography [DRA]) generates cone beam CT anatomical and vascular data in the DSA imaging setting, thus facilitating the correlation of 3D and 2D vasculature datasets for treatment planning. A robotic delivery system (Cyberknife; Accuray Inc, Sunnyvale, CA) uses frequent stereoscopic X-ray guidance to continuously localize the target and, in response to patient movements, redirect the beam during delivery. Based on these technologies, a frameless AVM radiosurgical treatment process can be envisaged that consists of the following steps: (1) a localization neuroangiography
INTRODUCTION Stereotactic radiosurgery is an important component of a multimodal arteriovenous malformation (AVM) treatment that historically uses a stereotactic frame rigidly fixed to a patient’s head. For treatment planning, imaging with a mode-specific localizer attached to the frame enables fusion of computed tomography (CT)/magnetic resonance (MR) and digital subtraction angiography (DSA) datasets, the last being essential for defining the nidus of some AVMs (1,2). Throughout radiosurgical treatment, attachment of the frame to the treatment device assures accurate localization and robust immobilization of the target. Reprint requests to: Dimitre Hristov, Ph.D., 875 Blake Wilbur Dr. Stanford, CA 94305-5847. Tel: (650) 498-7898; Fax: (650) 4984015; E-mail: [email protected] This work was presented in part at the 51st Annual Meeting of American Society for Therapeutic Radiology and Oncology, Nov 1–5, 2009, Chicago, IL.
Conflict of interest: Teri Moore and Marily Sarmiento are employees of Siemens Medical Solutions. John R. Adler is a shareholder of Accuray, Inc. Received Jan 7, 2010, and in revised form May 10, 2010. Accepted for publication May 14, 2010. 1232
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
procedure to acquire DRA and DSA datasets; (2) DSA-DRA (2D-3D) registration for spatial correlation of DSA and DRA image features; (3) AVM localization within the DRA image space by spatially linked navigation within the DRA and DSA datasets; (4) acquisition of a planning CT followed by DRA-CT registration and transfer of the DRA-localized target onto the CT image space; and (5) standard treatment planning and subsequent image-guided targeting utilizing Cyberknife. Colombo et al. (3) recently reported early results of Cyberknife radiosurgery using some of the methods proposed above. Significantly, frameless AVM localization based solely on DRA images and DRA/CT/MR registration was used (3). However, for AVM nidus definition, DRA by itself does not provide all the anatomic and physiologic information that comes with ‘‘time-stamped’’ DSAs. Therefore, for a comprehensive approach to frameless nidus localization, the incorporation of DSA is essential in many cases (1, 2). Thus, the purpose of the present project was to develop a process for frameless AVM localization consisting of steps 1 to 3, outlined above, and to evaluate the associated localization uncertainty.
METHODS AND MATERIALS C-arm DRA imaging A flat-panel C-arm imaging system (Axiom Artis; Siemens Medical Solutions, Inc, Malvern, PA) supporting various DRA modes was used. A DRA scan involves continuous C-arm rotation on a circular trajectory around a subject with concurrent acquisition of Xray projection images. The acquisition geometry for each DRA mode is calibrated first by scanning a PDS-2 calibration phantom (Siemens AG, Forchheim, Germany) with 108 spherical fiducial markers (Fig. 1A) embedded in a cylinder with a diameter of 145 mm. The axial extent of the fiducial pattern is about 145 mm. The known 3D marker positions in the coordinate system of the phantom and the 2D positions of the marker projections (Fig. 1B) in the coordinate systems of the individual CACT X-ray projections are used to derive 3 4 projection matrices which define a unique correspondence between a 3D point and its 2D projection (Fig. 1C) (4). These matrices are used in the DRA reconstruction for backprojection of the filtered projection images.
1233
The available DRA modes offer different compromises between acquisition time, image resolution, and contrast. For frameless AVM localization, we selected the ‘‘5s-1k DR 480 1.5 00.36’’ mode acquiring 126 projection images over 188 in 5 seconds for the reconstruction of pre- and postcontrast DRA volumes encompassing 256 256 222 voxels at 0.81-mm isotropic voxel size. While providing excellent visualization of contrast-filled cerebral vessels and high-contrast bony anatomy (Fig. A. 1, Supplementary Material), this mode also minimizes the probability of patient motion during image acquisition and maximizes the DRA reconstruction field-of-view to facilitate subsequent DRA-CT registration based on bony anatomy.
Spatial integrity of DRA and CT images For localization, DRA datasets need to accurately depict the spatial relationships between anatomical features. However, cone beam images reconstructed from a circular trajectory exhibit artifacts due to the lack of sufficient information for accurate 3D reconstruction ((5) and references therein). Furthermore, other potential geometric distortions may arise as a result of variances in the calibration and mechanical stability of the C-arm. The DRA spatial integrity was evaluated as follows. A DRA scan of the calibration phantom (Fig. 1A) was acquired and reconstructed. CT phantom images were also acquired according to a local radiosurgical simulation protocol: axial scanning; slice thickness, 1.25 mm; pixel size, 0.55 mm. The fiducial centers (Fig. 1A) within the DRA and the CT volumes were determined by intensity-based center-of-mass calculation within a volume-of-interest (VOI) encompassing each fiducial marker. The VOIs around the fiducials were segmented by a global, user-defined threshold. The lengths of segments connecting the centers of all possible fiducial pairs were calculated (Fig. 2). The mean and the standard deviation of the differences between the image-based segment lengths and the segment lengths derived from the phantom specifications served as metrics for the DRA and CT spatial integrity values.
Localization imaging protocol: description and reproducibility In order to facilitate the spatial correlation of DSA and DRA datasets, we introduced an imaging protocol consisting of DSA anteriorposterior (AP), DSA lateral (LAT), and DRA series with fixed couch position for all series. The parameters for these acquisitions were entered as preset programs in the imaging system to assure consistency and reproducibility of the acquisition geometries. The AP and LAT C-arm positions were selected by matching AP/LAT calibration
Fig. 1. (A) 3D rendering of the C-arm CT calibration phantom. (B) A projection image of the calibration phantom demonstrating the embedded fiducial markers. (C) 3D-2D point correspondence described by a projection matrix P.
1234
I. J. Radiation Oncology d Biology d Physics
Volume 79, Number 4, 2011
At the core of Angio-Localize is the procedure for registering DSA and DRA images detailed below.
DSA-DRA registration
Fig. 2. 3D rendering of the C-arm CT calibration phantom demonstrating the calculated centers of the fiducial markers. Sample segments connecting the centers of fiducial pairs are shown. phantom images to corresponding projections from the DRA series. Thus, provided that accuracy of the match is sufficient and consistently reproducible, the 3 4 matrices associated with the DRA projections could be used to link the DRA image space to the AP/LAT DSA series. We evaluated the geometric reproducibility of the DSA presets in two imaging sessions separated by 11 days. During each session, several AP and LAT views (Fig. 1B) of the calibration phantom were acquired. Prior to each series, the C-arm was parked and then placed in an arbitrary position and then instructed to move to the preset imaging position. The centers of the fiducial markers within the images (Fig. 1B) were determined by intensity-based centerof-mass calculation within a region-of-interest (ROI) encompassing each fiducial. The ROIs around the fiducials were segmented by a global, user-defined threshold. Overlapping fiducials clustered in the periphery of the radiographic projections (Fig. 1B) were excluded from the analysis, since their centers could not be reliably extracted by center-of-mass calculations. For the individual fiducial markers, the absolute deviations in position captured across a series of images were calculated. For consistency with the metrics used for DSA-DRA registration, the maximum values of the absolute deviations and population means of the absolute deviations were used as image-based metrics for geometry reproducibility.
Angio-Localize software We developed software (Angio-Localize) to enable integration, visualization, and exploration of various 3D and 2D datasets for frameless AVM localization. The application is based on a freely available medical visualization platform (6) and open source packages. Thus, the application can be deployed and enhanced by new users (http://www.stanford.edu/dhristov). Within the software, a typical workflow for frameless AVM localization is supported by five successively executed task-pad mini-applications (Table 1).
Alignment of the DSA (2D) and DRA (3D) datasets is a 2D-3D registration problem which requires the determination of a 3 4 ! projection matrix P such that for any point X within the DRA 3D ! coordinate space and its 2D projection x depicted in the DSA image ! (Fig. 1C), the equality lx ! x ¼ P X applies, where the scaling coef! ficient lx depends on X . Intensity-based (7, 8), feature-based (8–10), and hybrid (11, 12) techniques have been investigated for 2D–3D vascular registration. We have opted for a semiautomatic feature-based method based on segmented vascular structures (10). Figure 3 illustrates the 2D-3D registration process. Adjusting the viewing parameters used for DRA volume rendering, users interactively specify threshold levels for binary segmentation of 3D vasculature structures within a VOI (Fig. 3A). The centerline of the segmented vasculature is automatically extracted (13) and projected onto the DSA sequence (Fig. 3B). After reviewing the overlay of the projection with the DSA images, users select a particular DSA frame for 2D-3D registration and proceed by specifying threshold levels for segmenting DSA 2D vascular structures (Fig. 3C) within interactively selected ROI. The centerline of the segmented 2D vasculature is then automatically extracted (13) and converted to a distance map image (14) in which pixel values represent distance to the nearest point of the vascular centerline. Before proceeding with the automatic registration of the DRA-segmented vessel tree and the DSA-derived distance map image, users can exclude certain regions of the vascular tree from the registration process by defining ROIs around the projections of these regions on the distance map (Fig. 3D). The automatic 2D-3D registration is performed by solving a minimization problem with respect to the matrix parameters. The cost function is the mean distance between the projection of the DRAsegmented vessel tree centerline and the DSA-segmented vessel tree centerline. The cost is calculated by projecting the points from the 3D centerline onto the 2D distance map and accumulating the pixel values at the locations of the projected points. The matrix, P, can be decomposed as P = KC where K (intrinsic parameters) describes the projection geometry from the X-ray source to the flat-panel detector and C (extrinsic parameters) is a 4 4 affine transformation given as 2 3 tx 6 R ty 7 7; C ¼ ½R j T; T ¼ tx ; ty ; tz 0 : C¼6 4 tz 5 0 0 0 1 The extrinsic parameters describe the rotation, R, and translation, T, of the X-ray source in the DRA space. With square pixels of known dimensions, K can be written as 3 2 SID 0 u0 0 SID v0 0 5 K¼4 0 0 0 1 0 where SID is the source-to-imager distance, and u0 and n0 specify the image coordinates (in mm) of the piercing point of the normal from that of the X-ray source and the imaging plane. As a result of the 2D-3D registration, all 9 parameters (3 rotations, 3 translations, 3 intrinsic) need to be determined. However, small SID variations affect point projection coordinates only slightly. Furthermore, misestimation of u0 and n0 values can be somewhat compensated by adjustments in the X-ray source position. Thus, in our
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
1235
Table 1. Angio-Localize task pad mini-applications and their functionality Mini-application
Functionality
1. Patient selection 2. Image selection
Users create, select and manage records in a patient database. For a selected patient, users import 3D (CT, DRA, MR) and 2D (DSA) images into the database via DICOM interface. Subtraction of post- and pre-contrast datasets is performed. For a selected patient, users perform direct 3D-3D (DRA-CT, CT-MR) and 2D-3D (DSA-DRA) registration. Users review and approve registrations. Some datasets are further registered indirectly by concatenation of user-approved registrations. For a selected patient, users concurrently visualize and navigate registered 3D (CT, DRA, MR) and 2D (DSA) datasets. Users contour an AVM on a 3D dataset with visual feedback from the AVM display on other registered images. Users export AVM contours for treatment planning.
3. Registration 4. Localization 5. Export
implementation, the projection matrix is approximated as P = K0C0 [DR j DT]. In this expression K0 and C0 are fixed and estimated with information from the DICOM headers of the images, and [DR j DT] is a 4 x 4 affine matrix to account for small C-arm misalignments
(rotations, DR; and translations, DT) from the set positions and/or patient movements between image acquisitions. The objective function is thus minimized by a simplex method (15) with respect to the six parameters that determine [DR j DT].
Fig. 3. Steps in the DSA-DRA registration process. (A) Interactive segmentation of the DRA-defined 3D vasculature within a user-defined VOI. (B) Projection of points (visualized as markers) representing the centerline of the segmented 3D vasculature onto a DSA frame from the DSA series. (C) Interactive threshold segmentation of the DSA and overlay with the projected markers. A ROI (black outline) is defined for the subsequent processing steps. (D) Overlay of the projected markers with a distance map image derived from the segmented DSA. Regions (in black) are defined that contain marker outliers to be excluded from the automatic registration that minimizes the average image intensity across all markers. (E) Overlay of the projected markers with the distance transform image for visual inspection of the automatic registration results. (F) Overlay of the projected markers with user-selected DSA frame for visual inspection of the automatic registration results and interactive adjustment.
1236
I. J. Radiation Oncology d Biology d Physics
Volume 79, Number 4, 2011
Fig. 4. 2D-3D registration of DRA-segmented phantom fiducials to a radiograph. (A) Overlay of the projected fiducials and the radiograph that demonstrates initial misregistration. (B) Overlay of the projected fiducials with the radiographderived distance map used for the automatic 2D-3D registration. (C) As in panel B, after the registration. Fiducials excluded from the registration process are encompassed by user-defined ROIs (in black). (D) Overlay of the projected fiducials and the radiograph for visual confirmation of the registration results.
After the registration, P = K0C0[DR j DT] is used to project the 3D vessel tree onto the 2D distance map (Fig. 3E) and onto the DSA frames (Fig. 3F) to allow users to evaluate the registration quality. The cost and the number of centerline points used for its evaluation are reported. Furthermore, additional adjustment of the registration is possible by interactive manipulation of the parameters that determine [DR j DT] and visual evaluation of the alignment between the projection of the 3D vessel centerline and the DSA vasculature (Fig. 3F). We evaluated the inherent accuracy of the 2D-3D registration by registering a DRA scan of the calibration phantom to the AP and LAT views used for the evaluation of the imaging geometry reproducibility in session 2 (as described in Localization imaging protocol: description and reproducibility). An example is shown in Fig. 4. The centers of the calibration phantom fiducials (segmented as described in Spatial integrity of DRA and CT images; see also Fig. 1A and 2) were used as ‘‘segmented vasculature.’’ The residual cost at the end of the automatic registration was the 2D-3D registration metrics. This metric captures the quality of alignment between the fiducial projections and a distance map image. However, it does not consider the additional intrinsic registration uncertainty result-
ing from the discrete distance assignment when radiographs are converted into a distance map. Regardless of other factors, the discrete distance assignment places a lower limit of about 0.5 pixels on the achievable registration accuracy at the detector plane.
DSA-CT/MR registration and localization Within the Angio-Localize software, 3D-3D registration is used to link CT and MR volumes to pre-and postcontrast DRA datasets. The resulting transformations are concatenated with the DRA-toDSA registration matrices to generate perspective transformations that link CT and MR volumes to the DSAs. Once approved by a physician, these 3D-3D and 2D-3D transformations are used within the Localization mini-task for spatially correlated simultaneous navigation of 2D and 3D datasets (Fig. 5). In this process, as one follows the cerebral vessels on fused postcontrast DRA and contrastenhanced CT, the current 3D cursor location is projected on the registered LAT and AP DSA sequences. Similarly, AVM contours drawn on the 3D datasets are concurrently projected onto the DSA images for visual feedback. After the AVM segmentation, a DRA with an annotated AVM target is exported in DICOM format for planning.
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
1237
Fig. 5. AVM localization and contouring by correlated navigation of fused CT and postcontrast DRA datasets (top) and LAT and AP DSA dynamic sequences (bottom) within the Localization mini-task panel of Angio-Localize.
RESULTS Spatial integrity of DRA and CT images Table 2 summarizes the results from the evaluation of the DRA and CT spatial integrity. The systematic differences between image-derived and manufacturer-specified interfiducial distances were smaller than 0.05 mm (Table 2). Over the sample field of view defined by the markers, for the majority of the segments, the image-based interfiducial distances matched the manufacturer’s specifications within 0.5 mm (Table 2). The largest discrepancies were 0.6 mm for the DRA and 0.9 mm for the CT. Localization imaging protocol: geometry reproducibility Table 3 summarizes the results from the evaluation of the geometry reproducibility of the preset localization imaging Table 2. Differences between lengths of corresponding segments extracted from different datasets Parameter Segment length differences* No. of segments
DRA vs. phantom CT vs. phantom specifications specifications
DRA vs. CT
0.0 0.3
0.0 0.5
0.1 0.6
5,565
5,565
5,356
* Data are presented as means 2 standard deviation (mm).
modes. The mean deviations of the marker positions varied between 0.4 mm and 1.5 mm. Deviations as large as 3.2 mm were observed. These may have resulted either from intrinsic limits on the system’s mechanical stability or from operator variability in driving the system toward the set positions, since premature termination of the system movement is possible even if a set position is not reached. We did not investigate this further. Instead, we opted for the development of 2D-3D registration option in the AngioLocalize software to correct for C-arm misalignments from Table 3. Deviations observed in the image-based investigations of the geometric reproducibility of the preset C-arm modes* Session Mean absolute Maximum No. of images (preset) deviation (mm) deviation (mm) (no. of markers/image) 1 (AP) 1 (LAT) 2 (AP) 2 (LAT)
0.4 0.4 1.5 0.8
1.2 1.3 3.2 1.7
5 (51) 5 (44) 6 (48) 6 (44)
* Data show summary of the deviations observed in the imagebased investigations of the geometric reproducibility of the preset C-arm modes. The deviations are specified in the flat-panel detector plane. At the C-arm isocenter, these deviations are demagnified by a factor of 0.625.
1238
I. J. Radiation Oncology d Biology d Physics
Fig. 6. Lateral DSA image from left vertebral arteriogram reveals a tiny residual AVM nidus fed be the parietal occipital branch with early cortical draining vein to the superior sagittal sinus (arrow).
the set positions and/or patient movements between image acquisitions. DSA-DRA registration accuracy After the registration of the calibration phantom DRA scan to the six LAT radiographs in session 2, the residual cost function value calculated over 68 markers varied between 0.02 mm and 0.03 mm in the flat-panel detector plane. For the six AP radiographs, the residual cost function calculated over 58 markers varied between 0.02 mm and 0.04 mm in the detector plane. Thus, the discrete distance assignment in the distance map image is the dominant source of intrinsic registration uncertainty, about 0.5 pixels in the detector plane. With a pixel size of 0.308 mm and conservative radiographic magnification factor of 0.75 (typically 0.625 at C-arm isocenter), this registration uncertainty translates to inherent localization error of O3 0.5 0.308 0.75 mm = 0.2 mm or less within the DRA space. Exemplary clinical case The frameless technique was introduced for the treatment of a 23-year-old woman who required radiosurgery for a tiny postoperative residual AVM. This patient experienced one transient episode of left-sided weakness and numbness 3 years after undergoing surgical resection of a 3-cm-diameter, right posterior temporal parietal AVM, fed by the right posterior cerebral artery (PCA) and draining superficially into the superior sagittal sinus. Preradiosurgical angiography consisting of DRA and DSA acquisitions was performed with the patient under conscious sedation following the localization protocol described in Materials and Methods. During angiog-
Volume 79, Number 4, 2011
raphy, the patient’s head was immobilized by a Velcro strap. Both internal carotid and the dominant left vertebral arteries were selectively catheterized, and biplane digital subtraction angiographies were performed. The right posterior communicating artery was large with robust flow through the right posterior cerebral territory. A small, less than 1 cm, residual AVM nidus was identified, fed by a small parietal-occipital branch of the right PCA and with early drainage by way of a cortical vein and the posterior superior sagittal sinus, (Fig. 6). Subsequently, while immobilized in a head rest and custom face mask, the patient was imaged by using (1.25-mm) thin slice, contrast-enhanced axial CT using a pixel size of 0.55 mm. After approving the DSA-DRA registration (Fig. 7A), the treating neurosurgeon delineated the AVM nidus on the postcontrast DRA image set by concurrent linked navigation of the DRA and DSA images (Fig. 5). The postcontrast DRA with annotated AVM location was then exported to the treatment planning system (Multiplan version 2.2; Accuray Inc.). Treatment planning and delivery followed a standard process. Briefly, the DRA and the CT volume were automatically registered and fused to establish the position of the AVM within the planning CT (Fig. 7B). The commercial registration algorithm is based on normalized mutual information and a hierarchical optimization strategy, broadly similar to the technique described in the report by Studholme et al. (16, 17). A treatment plan was then generated to deliver 20 Gy in a single fraction to the nidus (Fig. 7B). Reference pairs of digitally reconstructed radiographs (DRRs) capturing different potential poses of the patient skull were generated by Cyberknife’s system. Prior to delivery, the patient was remobilized in the same custom accessories used with the planning CT. Then the target was brought to the treatment position by acquiring stereoscopic X-ray images, registering these to reference DRRs and adjusting the treatment table with the transformation parameters derived from registration. During the 90-min radiosurgery, the delivery system acquired an image pair approximately every 2 min. These images were concurrently correlated to the reference DRRs to redirect the radiation beam and track the target. DISCUSSION For intracranial targets that can be accurately localized within the coordinate space of high-resolution CT datasets, image-guided radiosurgery is a convenient alternative to frame-based approaches. Indeed, when assessed with the hidden target, Winston-Lutz test (18) in anthropomorphic phantoms, modern image-guided radiosurgery systems demonstrate targeting accuracy from 0.4 mm to 1.1 mm depending on the system and resolution of the CT scan used for target localization (19–21). Comparable targeting accuracy in the range of 0.3 mm to 1.3 mm has been reported with frame-based approaches (22, 23). Furthermore, frequent intrafractional imaging combined with skull-based tracking can reduce random targeting errors from patient motion
Frameless image-guided AVM radiosurgery d D. HRISTOV et al.
1239
Fig. 7. (A) Overlay of the projected segmented 3D vasculature with a frontal DSA and a lateral DSA for review and approval of the DSA-DRA registration. (B) Fused planning CT and postcontrast DRA images overlaid with the planned isodose distribution. The prescription dose is 20 Gy to the region encompassed by 80% (isodose line in green) of the maximum dose (25 Gy).
down to 1 mm (24), which is similar to the motion-induced error with frame-based radiosurgery (21). However, a comprehensive frameless radiosurgery technique for AVMs requires the integration of DSA datasets in the localization process (1, 2), which is the significance of the current work. We opted for a geometric, featurebased semiautomatic DSA-DRA registration, since the cost function indicates the potential localization error resulting from residual 2D-3D misalignment. While such a cost function was previously used (10), our implementation differs in two aspects. First, a faster optimization scheme is applied because the DSA and the DRA datasets are already approximately aligned as a result of the imaging protocol. Second, tools are provided for users to interactively remove outliers that can bias the registration. Such adjustments significantly improve the robustness of the registration process,
which is still contained within 5 to 10 min for an experienced user. The system-related localization error within the DRA space resulting from geometric distortions (Table 2) and automatic 2D-3D registration uncertainty (0.2 mm) is conservatively about 0.4 mm, thus comparable to a reported stereotactic targeting error of 0.4 0.1 mm based on biplanar radiographic imaging (22). Clinically, however, the localization error is determined by the accuracy of the physician-adjusted and -reviewed DSA-DRA and DRACT registrations. In our experience, during interactive 2D-3D registration, shifts as small as 0.2 mm in directions parallel to the imaging plane are readily identifiable through the inspection of the overlay of the projected vessel centerline and the DSA frames (Figs. 3F, 7A, and A. 2, Supplementary Material). On the other hand, even though the
1240
I. J. Radiation Oncology d Biology d Physics
DRA-CT registration is facilitated by the similar appearance of bony anatomy and the presence of some common vascular landmarks in the DRA and the contrast-enhanced CT (Fig. A.1), the local error is typically about 1 mm as evaluated from homologous landmarks in the images (Fig. A.1, Supplementary Material). Thus, the DRA-CT registration dominates the system-related uncertainty in the frameless AVM localization process. However, new angiography systems are capable of large field-of-view DRA imaging, and thus, DRA datasets can consistently capture the entire cranial anatomy. Thus, the DRA images may potentially be used on their own for the generation of DRRs as well as dose-calculation after appropriate electron density calibration.
Volume 79, Number 4, 2011
CONCLUSIONS In summary, the availability of DRA and image-guided delivery technology enables a comprehensive frameless radiosurgery process for AVMs. With this process, patient comfort is increased due to the elimination of pain or potential scar formation from placement of a stereotactic frame. Time pressure on the staff is reduced, since imaging, planning, and treatment sessions do not have to be completed in 1 day. The number of angiograms for a particular patient can also be potentially reduced by including DRA acquisition as part of the diagnostic imaging process, thereby eliminating the need for a subsequent imaging procedure used specifically for treatment planning.
REFERENCES 1. Zhang XQ, Shirato H, Aoyama H, et al. Clinical significance of 3D reconstruction of arteriovenous malformation using digital subtraction angiography and its modification with CT information in stereotactic radiosurgery. Int J Radiat Oncol Biol Phys 2003;57:1392–1399. 2. Hamm KD, Klisch J, Surber G, et al. Special aspects of diagnostic imaging for radiosurgery of arteriovenous malformations. Neurosurgery 2008;62:A44–52. 3. Colombo F, Cavedon C, Casentini L, et al. Early results of CyberKnife radiosurgery for arteriovenous malformations. J Neurosurg 2009 Oct;111(4):807–819. 4. Navab N, Bani-Hashemi A, Nadar MS, et al. 3D reconstruction from projection matrices in a C-arm based 3D-angiography system. Medical Image Computing and Computer-Assisted Intervention -. Miccai’98 1998;1496:119–129. 5. Bartolac S, Clackdoyle R, Noo F, et al. A local shift-variant Fourier model and experimental validation of circular conebeam computed tomography artifacts. Med Phys 2009;36: 500–512. 6. MeVisLab. Adevelopment environment for medical image processing and visualization, MeVis. Available at: http://www. MeVisLab.de. 2006. Accessed July 13, 2010. 7. Bullitt E, Liu A, Aylward SR, et al. Registration of 3D cerebral vessels with 2D digital angiograms: Clinical evaluation. Acad Radiol 1999;6:539–546. 8. McLaughlin RA, Hipwell J, Hawkes DJ, et al. A comparison of a similarity-based and a feature-based 2-D-3-D registration method for neurointerventional use. IEEE Trans Med Imaging 2005;24:1058–1066. 9. Groher M, Zikic D, Navab N. Deformable 2D-3D registration of vascular structures in a one view scenario. IEEE Trans Med Imaging 2009;28:847–860. 10. Florin C, Williams J, Khamene A, et al. Registration of 3D angiographic and X-ray images using sequential Monte Carlo sampling. Computer Vision for Biomedical Image Applications. Proceedings 2005;3765:427–436. 11. Jomier J, Bullitt E, Van Horn M, et al. 3D/2D model-to-image registration applied to TIPS surgery. In Larsen R, Nielsen M, Sporring Jon (editors). Medical Image Computing and Computer-Assisted Intervention - MICCAI 2006: 9th International Conference, Copenhagen, Denmark, October 2006 Proceedings. 2006;4191:662–669.
12. Vermandel M, Betrouni N, Pasquier D, et al. A 2D/3D matching based on a hybrid approach: improvement to the imaging flow for AVM radiosurgery. Conf Proc IEEE Eng Med Biol Soc 2005;3:3071–3073. 13. Selle D, Peitgen HO. Analysis of the morphology and structure of vessel systems using skeletonization. In: Chen C, Clough A, editors. Proc SPIE 2001;4321;. p. 271–281. 14. Fitzgibbon AW. Robust registration of 2D and 3D point sets. Image Vis Comput 2003;21:1145–1153. 15. Nelder JA, Mead R. A simplex-method for function minimization. Comp J 1965;7:308–313. 16. Studholme C, Hill DL, Hawkes DJ. Automated threedimensional registration of magnetic resonance and positron emission tomography brain images by multiresolution optimization of voxel similarity measures. Med Phys 1997;24:25–35. 17. Maurer CR Jr., West JB. Medical image registration using mutual information. In: Heilbrun MP, editor. Cyberknife radiosurgery: Practical guide 2. Sunnyvale, CA: The Cyberknife Society; 2006. p. 23–34. 18. Lutz W, Winston KR, Maleki N. A system for stereotactic radiosurgery with a linear accelerator. Int J Radiat Oncol Biol Phys 1988;14:373–381. 19. Antypas C, Pantelis E. Performance evaluation of a CyberKnife G4 image-guided robotic stereotactic radiosurgery system. Phys Med Biol 2008;53:4697–4718. 20. Lamba M, Breneman JC, Warnick RE. Evaluation of imageguided positioning for frameless intracranial radiosurgery. Int J Radiat Oncol Biol Phys 2009;74:913–919. 21. Solberg TD, Medin PM, Mullins J, et al. Quality assurance of immobilization and target localization systems for frameless stereotactic cranial and extracranial hypofractionated radiotherapy. Int J Radiat Oncol Biol Phys 2008;71:S131–S135. 22. Yeung D, Palta J, Fontanesi J, et al. Systematic analysis of errors in target localization and treatment delivery in stereotactic radiosurgery (SRS). Int J Radiat Oncol Biol Phys 1994;28:493– 498. 23. Rahimian J, Chen JC, Rao AA, et al. Geometrical accuracy of the Novalis stereotactic radiosurgery system for trigeminal neuralgia. J Neurosurg 2004;101(Suppl 3):351–355. 24. Murphy MJ. Intrafraction geometric uncertainties in frameless image-guided radiosurgery. Int J Radiat Oncol Biol Phys 2009;73:1364–1368.