Digital image remapping

Abstract

The present invention relates to production of 2D digital images suitable for use in medical imaging. The invention particularly relates to remapping X-ray images taken from a first viewpoint so that they present the same image as seen from a second viewpoint. Remapping is achieved by registering separate 2D images taken from the first and second viewpoints of an area with a 3D scan volume of the same region to ascertain their relative viewpoints with respect to the 3D scan volume. The image taken with respect to the first viewpoint is then remapped to yield the image as seen from the second viewpoint.

Claims

1. An image generation method comprising: a) obtaining single first and single second 2D images through an object to be imaged, wherein the images are from different viewpoints with respect to the object; b) providing a 3D image data set of the object to be imaged; c) defining a remapping surface within the 3D image data set; d) registering the first and second 2D images with the 3D image data set and determining a viewpoint of the first and second 2D images with respect to the 3D image data set; and e) remapping pixels of the first 2D image to generate a remapped 2D image from the viewpoint of the second 2D image by: i. back projecting rays from the pixels of the first 2D image and determining the points of intersection of the back projected rays with the remapping surface; and ii. remapping the pixels of the first 2D image corresponding to the points of intersection to generate the 2D remapped image, whereby the pixels of the first 2D image are remapped according to the direction of rays projected from the viewpoint of the second 2D image via the intersection points corresponding to each pixel.

2. The image generation method of claim 1, comprising: f) obtaining a third 2D image through an object to be imaged, wherein the third image shares the viewpoint of the first image and is aligned with the first image; g) providing a 3D image data set of the object to be imaged; h) defining a remapping surface within the 3D image data set; i) registering the first and second 2D images with the 3D image data set and determining the viewpoint of the first and second 2D images with respect to the 3D image data set; j) remapping pixels of the third 2D image to generate the remapped 2D image from the viewpoint of the second 2D image by: i. determining which pixels of the third 2D image correspond to the pixels of the first 2D image; ii. back projecting rays from the pixels of the first 2D image and determining the points of intersection of the back projected rays with the remapping surface; and iii. remapping the pixels of the third 2D image that correspond with the pixels of the first 2D image to generate the 2D remapped image, whereby the pixels are remapped according to the direction of rays projected from the viewpoint of the second 2D image via the intersection points corresponding to each pixel.

3. The method according to claim 1, wherein the first or second 2D image is a fluoroscopy image.

4. The method according claim 1, wherein the 3D image data set is obtained from a computerised tomography (CT) or magnetic resonance (MR) or cone beam computerised tomography (CBCT) scan.

5. The method according to claim 1, wherein the remapping surface is any of: i. angled planar or multi-planar; ii. curved; and/or iii. discontinuous; or any combination thereof.

6. The method according to claim 1, wherein the remapping surface is shaped to follow a feature of the object to be imaged.

7. The method according to claim 6, wherein the feature of the object to be imaged within the 3D image data set of the object to be imaged is enhanced by the use of contrast medium.

8. The method according to claim 1, wherein the origin of the pixels is marked.

9. The method according to claim 1 wherein the object is a patient and wherein the image generated by the remapping is used to carry out a surgical procedure on the patient.

10. The method of claim 1, further comprising administering contrast medium to the object to be imaged and thus defining a remapping surface within a 3D image data set in relation to the location of the contrast medium.

11. An imaging system, comprising: a 2D imaging system arranged in use to obtain 2D images; and a processor, arranged in use to: a) obtain single first and single second 2D images through an object to be imaged, wherein the images are from different viewpoints with respect to the object; b) obtain a 3D image data set of the object to be imaged; c) define a remapping surface within the 3D image data set; d) register the first and second 2D images with the 3D image data set and determine the viewpoint of the first and second 2D images with respect to the 3D image data set; and e) remap pixels of the first 2D image to generate a remapped 2D image from a viewpoint of the second 2D image by: i. back projecting rays from the pixels of the first 2D image and determining the points of intersection of the back projected rays with the remapping surface; and ii. remapping the pixels of the first 2D image corresponding to the points of intersection to generate the 2D remapped image, whereby the pixels of the first 2D image are remapped according to the direction of rays projected from the viewpoint of the second 2D image via the intersection points corresponding to each pixel.

12. The imaging system according to claim 11, wherein the 2D imaging system is arranged to obtain 2D images to be registered with a 3D image data set; and further comprising a processor, arranged to: f) obtain single first and single second 2D images through an object to be imaged, wherein the images are from different viewpoints with respect to the object; g) obtain a third 2D image through an object to be imaged, wherein the third image shares the viewpoint of the first image and is aligned with the first image; h) define a remapping surface within the 3D image data set; i) register the first and second 2D images with the 3D image data set and determining the viewpoint of the first and second 2D images with respect to the 3D image data set; j) remap pixels of the third 2D image to generate the remapped 2D image from the viewpoint of the second 2D image by: i. determining which pixels of the third 2D image correspond to the pixels of the first 2D image; ii. back projecting rays from the pixels of the first 2D image and determining the points of intersection of the back projected rays with the remapping surface; and iii. remapping the pixels of the third 2D image that correspond with the pixels of the first 2D image to generate the 2D remapped image, whereby the pixels are remapped according to the direction of rays projected from the viewpoint of the second 2D image via the intersection points corresponding to each pixel.

Description

(1) The invention is now illustrated with reference to the following specific examples and the accompanying drawings which show:

(2) FIG. 1 Images illustrating basic digital subtraction angiography. The mask image (a) is subtracted pixel-by-pixel from all individual frames in the angiography screening (b.sub.1˜b.sub.n). A DSA image (c) is then generated from the subtracted frames using the maximum intensity projection method.

(3) FIG. 2 Illustration of the problem of remapping projection data. (a) The 2D circle in view 1 cannot be remapped from view 1 to view 2 without knowing it's 3D position (i.e. grey circle) along the ray path for view 1. (b) If a remapping surface is known, the square, triangle and circle seen in view 1 can be correctly remapped from view 1 to view 2.

(4) FIG. 3 The 10 degrees of freedom involved in perspective projection transformation. A preoperative CT volume is registered with the intraoperative image using a 2D-3D registration algorithm. c.sub.s and I.sub.s mark the positions of the interception between the ray projected from the X-ray source into the imaging plane. θ.sub.x, θ.sub.y and θ.sub.z represent the imaged object orientation, while X, Y and Z represent its position with respect to the fluoroscopy set coordinate system X.sub.3D.

(5) FIG. 4 Flow diagram showing how the 2D-3D registration algorithm enables DSA remapping. Input images are shown at the top: (a) DSA mask; (b) DSA image; (c) new fluoroscopy image acquired after the C-arm was moved to a new view; and (d) a preoperative CT volume comprising the remapping surface. Middle and bottom shows the 2D-3D registration which enables calculation of view directions and positioning of patient-vasculature-specific remapping surface (e).

(6) FIG. 5 DSA remapping process shown in detail for one pixel in the DSA image: I.sub.DSA(u,v). A ray is back projected from I.sub.DSA(u,v) into the remapping surface using M.sub.DSA. The 3D interception position CT (x,y, z).sub.int is then projected into the fluoroscopy image using the transformation M.sub.FL to acquire the 2D interception position in the fluoroscopy image I.sub.FL (u,v). Finally, the intensity at I.sub.DSA(u,v) is remapped onto I.sub.FL (u,v).

(7) FIG. 6 DSA remapping Type 1 error E.sub.1, as a function of the feature's thickness and changes in ray paths. Both (a) and (b) show images of the aorta acquired from two different views with the upward-pointing arrow intersecting a different part of the aorta than the downward-pointing arrow (indicated by solid lines), thus causing a type 1 error. In addition, (b) shows that when a feature's thickness in (b) is much larger than in (a) the solid line originating from the upward-pointing arrow is further away from the solid line originating from the downward-pointing arrow when compared to (a).

(8) FIG. 7 DSA remapping Type 2 error E.sub.2 relation with (a) the 2D-3D registration errors E.sub.reg, and (b) the intraoperative deformation (D.sub.ef). In both cases two images of the aorta are acquired from different views. The downward-pointing arrow intersects the CT volume at the wrong depth position causing E.sub.2>0 in both (a) and (b).

(9) FIG. 8 A surface used for DSA remapping from different anatomical views of an aorta: (a) anterior-posterior, (b) lateral and (c) posterior-anterior. The surface is defined along the aorta using thin-plate-spline interpolation.

(10) FIG. 9 Illustration of a validation method used. (a) Fluoroscopy image with a dot marking a guide-wire. (b) Remapped DSA image wherein a dot marks the renal ostium. (c) Rays are back projected from the centres of each of the dots in the overlay image into the remapping surface using the transformation I.sub.FL, and the error is calculated in millimetres in CT.sub.3D.

(11) FIG. 10 Representative results from patients 2, 4 and 8 respectively. (a) I.sub.DSA, (b) I.sub.FL, (c) I.sub.DSA(rem) and (d) I.sub.DSA(rem) overlaid onto I.sub.FL in red, with the renal ostia marked with black crosses (+) in I.sub.DSA(rem) and white crosses () in I.sub.FL.

(12) As noted above in relation to FIG. 2, FIG. 2a shows that remapping a first DSA image to the second viewing direction cannot be accurately achieved using only knowledge of the relative directions of views 1 and 2.

(13) Additional knowledge is required: information on the depth position (i.e. the distance along the ray path) of anatomical features inside the patient. In FIG. 2a, the grey circle inside the patient is projected into the detector along the dotted line using the first view direction. However, when the source is moved to the second view direction, the circle seen in view 1 cannot be directly remapped to view 2.

(14) To perform such a remapping extra information on the grey circle's position inside the patient is required. This is because the circle seen in view 1 might be projected from any point along the blue line intersecting the patient, such as the points marked with +, which if used for remapping along the ray lines for view 2, result in multiple possible locations for the circle in view 2.

(15) However, if all the 2D image information can be projected back to a single accurately known 2D surface, such as the one depicted in FIG. 2b, then the image can be remapped accurately for view 2. In FIG. 2b, the projected square, triangle and circle of view 1 can be correctly remapped to view 2 using the positional information provided by the remapping surface intersecting these features inside the patient.

(16) Therefore it is possible to define a remapping surface inside the preoperative CT volume and then match this surface to the patient using 2D-3D registration during an operation or intervention. This surface is then used to remap a DSA image to a new view direction.

(17) As the depth information becomes more three-dimensional (i.e. does not just originate from a single 2D remapping surface), and as errors arise in positioning the 2D surface, then errors arise in the remapping process. Consequently, the ability to define such a surface accurately is essential for a correct perspective projection remapping.

(18) Using a 2D-3D Registration Algorithm to Facilitate DSA Remapping

(19) The present disclosure utilises the 2D-3D registration algorithm noted above to facilitate DSA remapping. FIG. 4 demonstrates how the 2D-3D registration algorithm enables DSA remapping. This begins at the top with the input images: (a) a DSA mask and (b) a DSA image (I.sub.DSA) produced from a posterior-anterior view; (c) a new fluoroscopy image (I.sub.FL) acquired after the C-arm is moved; and (d) the preoperative CT scan with the remapping surface defined inside. Images (a), (c) and (d) are input into the 2D-3D registration algorithm which calculates the 2D-3D transformations between the CT scan (x,y,z,1).sup.T and both the DSA image (u,v,1).sub.DSA.sup.T and the new fluoroscopy image (u,v,1).sub.FL.sup.T i.e.:
M.sub.DSA(x,y,z,1).sup.T=λ(u,v,1).sub.DSA.sup.T  (3.1)
M.sub.FL(x,y,z,1).sup.T=λ(u,v,1).sub.FL.sup.T  (3.2)

(20) The three boxes in FIG. 4 show the subsequent stages to the 2D-3D registration which provide the necessary information to carry out DSA remapping. The Roman numerals labelling each box correspond to the following processes: I. Calculate DSA view position: the transformation M.sub.DSA to position I.sub.DSA in relation with CT.sub.3D is determined using the DSA mask. Both the DSA mask and DSA image have the same transformation as they are acquired in a single angiography screening from the same view direction. II. Calculate new fluoroscopy view position: the transformation M.sub.FL to position I.sub.FL in relation with CT.sub.30 is determined. III. The transformations M.sub.DSA and M.sub.FL can position the preoperatively defined remapping surface inside the CT volume with respect to both I.sub.DSA and I.sub.FL, enabling remapping to occur on a patient-vasculature-specific surface.

(21) DSA Remapping

(22) Thus the 2D-3D registration enables DSA remapping and is detailed in the below steps (I, II, III, IV, V) with reference to FIG. 5; the Roman numerals in FIG. 5 correspond to the following steps: I. Segmenting a remapping surface from the preoperative CT volume. The surface should contain the blood vessels of clinical interest to be remapped. II. Registering the CT volume with both images, I.sub.DSA and I.sub.FL, using the 2D-3D registration algorithm such as that described herein. The registration allows spatial positioning of I.sub.DSA and I.sub.FL in relation to the segmented CT remapping surface as was described above with reference to boxes I and II of FIG. 4. III. Back projecting rays from each of the DSA image pixels I.sub.DSA(u,v) using M.sub.DSA, and calculating the 3D positions in CT.sub.3D where the rays intercept the remapping surface (i.e. CT (x,y,z).sub.int). IV. Projecting rays from the 3D interception positions CT (x,y,z).sub.int to I.sub.FL to acquire the 2D interception position in I.sub.2D (i.e. I.sub.FL(u,v)). This is done using the transformation M.sub.FL. V. Finally, the intensity at each DSA image pixel I.sub.DSA(u,v) is remapped onto the 2D interception position I.sub.FL(u,v) corresponding to the same pixel. This automatically produces a remapped DSA image corresponding to the current fluoroscopy view.

(23) DSA Remapping Errors

(24) As discussed in above, remapping a projection image into a new view requires knowledge of the depth position of the anatomical features. As noted herein, such knowledge can be provided by defining a remapping surface inside the CT volume to intersect features of interest. This assumes that features lie on a single 2D surface, and that the surface can be positioned correctly. However, when these assumptions are violated three types of errors arise: I. We define type 1 error E.sub.1 to occur as a result of the thickness of the feature being remapped (i.e. feature's size along the z direction in CT.sub.3D. E.sub.1 is a function of the feature's thickness and changes in the ray paths (ΔrayPath) along that feature when the C-arm is moved to a new view direction, such as:

(25) $E_1=f\left(\mathrm{thickness},\Delta \mathrm{rayPath}\right)\{\begin{array}{c}{E}_1=0,\mathrm{if}\mathrm{features}\mathrm{lie}\mathrm{completely}\mathrm{on}\mathrm{the}\mathrm{remapping}\mathrm{surface}.\\ E_1>0,\mathrm{otherwise}\end{array}.$

(26) When the entire feature lies on the remapping surface, then, the ray paths along that feature are similar from any view direction and E.sub.1. However, when the feature's thickness increases, the changes in the ray paths increase when the C-arm is moved, and thus, E.sub.1 increases as seen in FIG. 6.

(27) In FIGS. 6(a) and (b), two images of an aneurysmal aorta are acquired from two different view directions. In both cases, the projected ray to view 2 (upward-pointing arrow) does not intersect the same part of the aorta as the projected ray from view 1 (downward-pointing arrow). This causes type 1 error in the features' position between the remapped image from view 1 and the new image from view 2 when overlaid. In addition, in (b), the intersection of the upward-pointing arrow with the aorta (solid line) is further away from the intersection of the downward-pointing arrow (solid line) when compared to (a), this is because the feature's thickness in (b) is much larger than in (b). II. We define type 2 error E.sub.2 to be due to errors in positioning the remapping surface. E.sub.2 is a function of the 2D-3D registration error (E.sub.reg [9]), and the intraoperative deformation (D.sub.ef [8]), such as:

(28) $E_2=f\left(E_{\mathrm{reg}},D_{\mathrm{ef}}\right)\{\begin{array}{c}{E}_2=0,\mathrm{if}{E}_{\mathrm{reg}}=0\mathrm{and}{D}_{\mathrm{ef}}=0.\\ E_2>0,\mathrm{otherwise}\end{array}.$

(29) E.sub.reg results from misaligning CT.sub.3D with X.sub.3D (i.e. errors in R(θx,θy,θz) and T(X,Y,Z). The translation error along the Z axis (see FIG. 3) is the largest error observed when compared to the other translation and rotation errors (5 mm vs. 0.5 mm respectively [9]). Therefore, errors in positioning the remapping surface along the Z axis may occur as shown in FIG. 7(a). In FIG. 7(a), where no deformation occurs (D.sub.ef=0, the downward-pointing arrow intersects the CT volume at the wrong depth position because of the Z translation error causing E.sub.2 error.

(30) D.sub.ef is a common issue in all methods that employ preoperative anatomy for overlay. D.sub.ef might occur because of the movement of the stiff wires and delivery systems inside the aorta during intervention (<10 mm [8]). This might cause errors in the position of the remapping surface as illustrated in FIG. 7(b). In FIG. 7(b), where no registration error occurs E.sub.reg, the downward-pointing arrow intersects the CT volume at the wrong depth position because of the intraoperative deformation of the remapping surface causing E.sub.2 error. III. We define type 3 error E.sub.3 to be due to non-rigid movement of features of interest (relative to the vertebrae on which registration is based) between the time the DSA image I.sub.DSA was acquired and the time the new fluoroscopy image (I.sub.FL) was acquired. E.sub.3 is a function of the different stages during an intervention which exhibit different amounts of intraoperative deformation depending on the type of the interventional devices present.

(31) E.sub.3 is a minimum when I.sub.DSA and I.sub.FL are acquired during the same stage of the intervention as the amount of deformation should be similar for both images. However, if I.sub.DSA and I.sub.FL are acquired during different stages, then E.sub.3 increases as the two images experience different amounts of deformation. In addition, the delivery device with the undeployed stent-grafts is the main cause of deformation. Therefore, if I.sub.DSA is acquired when the delivery device is present and I.sub.FL when it was not, E.sub.3 is a maximum.

EXAMPLES

(32) Experiments were carried out using data from 9 patients who underwent elective EVAR in St Thomas' hospital (London, UK). Data was processed offline, i.e. not during the procedure and was approved by the National Research Ethics Service with informed patient consent. The 2D-3D registration was performed on a computer with two NVidia GTX 690 graphic cards with each card containing two GPUs. A single 2D-3D registration was completed in 1.25 sec, and the remapping software took around 1 sec. Thus, the entire remapping process time was around 3.5 sec for each fluoroscopy image.

(33) Each dataset had a preoperative diagnostic CT scan, acquired on a variety of machines depending on the referring hospital, with voxel sizes ranging from 0.683×0.683×0.7 mm.sup.3 to 1×1×1 mm.sup.3. Each dataset also had a number of intraoperative images (fluoroscopy screening, angiography screening and DSAs) acquired on a Siemens FP20 system with a low frame rate ranging from 2 fps to 7 fps.

(34) For each patient, the aorta was segmented from the preoperative CT volume using a semi-automatic method in ITK-SNAP [11]. The remapping surface was then defined by picking points along the midline of the aorta, iliac and renal arteries and then producing a surface using thin-plate-spline interpolation [12]. An example of such a remapping surface can be seen in FIG. 8 from different views: (a) anterior-posterior, (b) lateral and (c) posterior-anterior. This surface was chosen to remap blood vessels of interest, i.e. the aorta, and renal and iliac arteries.

(35) For each dataset, a DSA image produced from an anterior-posterior view at an early stage of the intervention was chosen to be remapped. These DSA images show the delivery devices with the undeployed stent-grafts as well as vasculature. DSA remapping, as described herein, was then carried out to remap the chosen DSA image to a number of fluoroscopy images acquired at different stages of the intervention after C-arm movement (but approximately from anterior-posterior views).

(36) Validation Experiments

(37) Validation images were chosen for each dataset which clearly showed the position of the renal arteries, either by the position of a guide-wire or a stent-graft, or by the use of ICM. Overlay accuracy was then calculated at a clinically relevant position: the renal ostium in both the fluoroscopy image I.sub.FL and the remapped DSA image I.sub.DSA(rem), as shown in FIG. 9. In FIG. 9, where a guide-wire is used, an error value of zero was recorded if the wire in I.sub.FL (light-toned dot in (a)) went through the I.sub.DSA(rem) renal ostium (dark-toned dot in (b)). Otherwise, the error value was calculated as described in FIG. 9(c). In FIG. 9(c), rays are back projected from the overlay image into the remapping surface using the transformation I.sub.FL. The distance (i.e. error) between the two points of interception with the surface is then calculated in mm in CT.sub.3D using the formula √{square root over ((x.sub.1−x.sub.2).sup.2+(y.sub.1−y.sub.2).sup.2+(z.sub.1−z.sub.2).sup.2)}. This method allows the error to be calculated in real anatomical distance (i.e. mm) not a projected error (i.e. pixels).

(38) In the case where a stent-graft or an ICM was used, the middle-points of the renal ostia were located in both I.sub.FL and I.sub.DSA(rem), and the same method of back projection was used to calculate the error in millimetres. For each dataset and in all cases (i.e. wire/stent/ICM), locations of the renal ostia were located visually by two observers independently. Two sets of errors were then calculated and averaged for each dataset.

(39) FIG. 10 shows results from patients 2, 4 and 8 respectively: (a) the DSA image (I.sub.DSA); (b) the fluoroscopy image (I.sub.FL); (c) the remapped DSA image (I.sub.DSA(rem); and (d) I.sub.DSA(rem) overlaid onto I.sub.FL in red, with the renal ostia marked with black crosses (+) in I.sub.DSA(rem) and white crosses () in I.sub.FL. Presented patients were chosen to have remapping errors which covered the full range of observed average errors: 0.82 mm, 2.92 mm and 5.52 mm, respectively.

(40) For each patient, the remapping accuracies, as described above, were calculated for all images and averaged. Table 1 lists the number of DSA remappings for each patient, and the maximum and averaged remapping errors. The overall number of remappings performed and averaged error are also presented. Numerical results showed an overall error average of 2.50 mm over 41 remapped images, with 1 case scoring zero error and 6 other cases scoring averaged errors <3 mm. For 2 patients, larger averaged errors (>4 mm) were observed. In 5 patients, large maximum errors (>4 mm) were observed; patient 8 scored the highest maximum (11.57 mm) and averaged (5.52 mm) errors.

(41) Results presented in Table 1 show an averaged remapping error of 2.50 mm over 41 remappings performed. Error variations across different remappings for the same patient were observed. This can be explained by the fact that remappings were performed at different stages of the procedure for each patient (as noted above). Thus, the aorta experienced different amounts of intraoperative deformation depending on the type of interventional devices present. When overlaying the segmented aorta into the remapped DSA image with the highest overlay error (i.e. 11.57 mm for patient 8), this case was found to have the biggest intraoperative deformation compared to the other DSA remappings for the same patient, which explains the larger maximum error reported for patient 8.

(42) TABLE-US-00001 TABLE 1 The number of DSA remappings performed, the maximum error and the averaged remapping error in millimetres (mm) for each patient. Number of Maximum Error Averaged Error Remappings (mm) (mm) Patient 1 5 3.39 1.79 Patient 2 7 1.34 0.82 Patient 3 5 4.41 2.46 Patient 4 4 4.55 2.92 Patient 5 3 0 0 Patient 6 2 3.1 1.77 Patient 7 6 4.64 2.61 Patient 8 6 11.57 5.52 Patient 9 3 6.19 4.58 Overall 41 11.57 2.5

(43) Warping the preoperative aorta to match the intraoperative scene before segmenting the remapping surface can also be investigated by employing a non-rigid 2D-3D image registration algorithm (e.g. [13]) to account for the intraoperative deformation.

(44) Accordingly, the invention provides methods and systems for production of 2D digital images suitable for use in medical imaging.

REFERENCES

(45) [1] Miller, D., “Interventional fluoroscopy: reducing radiation risks for patients and staff,” J. Vasc. Interv. Radiol. 20(7), 274-274 (2009).

(46) [2] Patel, A., Gallacher, D., Dourado, R., Lyons, O., Smith, A., Zayed, H., Waltham, M., Sabharwal, T., Bell, R., Carrell, T., et al., “Occupational radiation exposure during endovascular aortic procedures,” Eur. J. Vasc. Endovasc. 46(4), 424-430 (2013).

(47) [3] McCullough, P. A., “Contrast-induced acute kidney injury,” J. Am. Coll. Cardiol. 51(15), 1419-1428 (2008).

(48) [4] Nash, K., Hafeez, A., and Hou, S., “Hospital-acquired renal insufficiency,” Am. J. Kidney. Dis. 39(5), 930-936 (2002).

(49) [5] Seeliger, E., Sendeski, M., Rihal, C. and Persson, P., “Contrast-induced kidney injury: mechanisms, risk factors, and prevention,” Eur. Heart. J. 33(16), 2007-2015 (2012).

(50) [6] Bicknell, C., “Occupational radiation exposure and the vascular interventionalist,” Eur. J. Vasc. 46, 431 (2013).

(51) [7] Markelj, P., Toma{hacek over ( )}evi{hacek over ( )}z c, D., Likar, B. and Pernu{hacek over ( )}s, F., “A review of 3D/2D registration methods for image-guided interventions,” Med. Image. Anal. 16(3), 642-661 (2012).

(52) [8] Carrell, T., Modarai, B., Brown, J. and Penney, G., “Feasibility and limitations of an automated 2D-3D rigid image registration system for complex endovascular aortic procedures,” J. Endovasc. Ther. 17(4), 527-533 (2010).

(53) [9] Penney, G., Weese, J., Little, J., Desmedt, P., Hill, D., et al., “A comparison of similarity measures for use in 2D/3D medical image registration,” IEEE Trans. Med. Imag. 17(4), 586-595 (1998).

(54) [10] Penney, G., Varnavas, A., Dastur, N. and Carrell, T., “An image-guided surgery system to aid endovascular treatment of complex aortic aneurysms: description and initial clinical experience,” in [IPCAI], 6689, 13-24 (2011).

(55) [11] Yushkevich, P. A., Piven, J., Hazlett, H. C., Smith, R. G., Ho, S., Gee, J. C. and Gerig, G., “User guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability,” Neuroimage 31(3), 1116-1128 (2006).

(56) [12] Rohr, K., Fornefett, M. and Stiehl, H. S., “Approximating thin-plate splines for elastic registration: Integration of landmark errors and orientation attributes,” in [Inf. Process. Med. Imaging], 1613, 252-265, Springer (1999).

(57) [13] Guyot, A., Varnavas, A., Carrell, T. and Penney, G., “Non-rigid 2D-3D registration using anisotropic error ellipsoids to account for projection uncertainties during aortic surgery,” in [MICCAI], 8151, 179-186, Springer (2013).