The template matching method is a technique to determine corresponding points by evaluating the coincidence between a template image and region of interest ...
Int J CARS (2013) 8 (Suppl 1):S115–S118
Fig. 1 Configuration and workflow of our 3D heart surface motion measurement system apply the template matching method. The template matching method is a technique to determine corresponding points by evaluating the coincidence between a template image and region of interest (ROI). In this study, because the contrast of endoscopic images of the heart surface was low, it was hard to acquire features on the heart surface image; thus, we emphasize the features on the ROI image. To emphasize the features, the ROI image is smoothed by its histogram, and the edges of the smoothed ROI image are extracted by difference of Gaussian filtering. The ROI image is then smoothed by its histogram and a median filter again. The tracking points and template images are manually selected on the ROI image, and the ROI image is used to track a template image in the next frame. The template image is updated by evaluating the coincidence between the template image and ROI image during tracking. If the maximum coincidence is lower than 50 %, a template image is initialized. If the coincidence is lower than 50 % in the next frame after initialization, the template image is initialized using a template image of the first frame. To measure the heart surface motion, three tracking points are used on each endoscope image; the linear and rotational motions of the rigid plane consisting of these three points are measured. The linear motion is measured using the centroid motion of the plane, and the rotational motion is measured using the angle of the plane against the origin of the integrated coordinate. Figure 2 shows the overview of the experimental environment. To evaluate the system, we used a silicon heart model and electromagnetic tracking system (Aurora, Northern Digital, Inc.). The heart model was moved manually to act like a beating heart. An electromagnetic tracking sensor was attached to the surface of the heart model, and the model surface motion was measured in six degrees of freedom. The measured value was used as the ground truth of the heart surface motion. To evaluate our measurement system,
S117 registration between Polaris and Aurora coordinate systems was performed using four fiducial points. The fiducial registration error of this experimental setup was approximately 0.36 mm. The 3D heart surface motion was evaluated based on the relative displacements of the electromagnetic tracking sensor and the measured centroid and angle motions of the rigid plane consisting of the three feature points tracked with our endoscope-based measurement system. The resolutions of the input endoscope images, ROI images, and template images were 6400 480, 1410 141, and 810 81 pixels, respectively. The initial positions of the three tracking points on each endoscopic image were manually selected. The tracking was performed over 300 frames. Results The experimental results showed that the average errors of the linear motion were 2.31 ± 2.18, 4.56 ± 2.51, and 1.65 ± 2.10 mm for the x, y, and z coordinates, respectively, and 5.87 ± 3.15 mm in 3D space (n = 300). The average errors of the rotational motion were 2.71 ± 3.23°, 3.58 ± 5.97°, and 8.23 ± 37.84° for the x, y, and z coordinates, respectively (n = 300). The average computing time from image acquisition to location output was 24.31 ± 10.15 ms. Conclusion We developed a 3D heart surface motion measurement system with two endoscopes using the template matching method. Measurement errors include the calibration error of endoscope images, registration error between the Polaris and Aurora systems, and tracking error of the template matching method. The calibration error has a potentially large effect, in particular, because the registration error was low at approximately 0.36 mm, and 2D point tracking of each image was performed accurately. In this experiment, we set the two endoscopes close together because we had to take images on the same surface area for these two endoscopes. Therefore, the stereo camera calibration error greatly influenced the position error accuracy of the tracked points in 3D space. To overcome this problem, we have to use an endoscope with a wide viewing angle and high resolution. Acknowledgement This research is partly supported by the Fund for the Improvement of Research Environment for Young Researchers and KAKENHI grant by MEXT (23680056, 22650115). Reference [1] Hagihara D, Takahashi T, Sato I, Nakamura R (2012) Tracking of heart surface motion for myocardial cell sheet implanation. Int J CARS 7(suppl 1): S386–S387
Dynamic overlay of an intra-operative aortic root model onto X-ray fluoroscopy sequences M.E. Karar1, O. Burgert2 1 University of Menofia, Faculty of Electronic Engineering, Menouf, Egypt 2 Reutlingen University, Dept. of Medical Computer Engineering, Reutlingen, Germany Keywords Aortic root Minimally invasive surgery X-ray fluoroscopy Image-based tracking
Fig. 2 The experimental environment
Purpose Image-guided minimally invasive trans-catheter aortic valve implantation (TAVI) is still limited to static overlay of projected 3-D aortic root model onto live 2-D X-ray fluoroscopic images. That causes misalignments between the static overlaid aortic model and the underlying anatomical structures because of heart beating and respiratory motion. Also, X-ray fluoroscopy guidance has limitations to guide the TAVI procedure. The contrast of X-ray images is generally low. Only contrast agent is injected to visualize the aortic root, valve annulus, and coronary ostia, but an excessive usage of contrast agent may cause renal insufficiency in high-risk patients [1].
123
S118 In order to intra-operatively assist visualization of the aortic root in X-ray images with and/or without contrast agent, we present a method to continuously update the overlaid aortic root model from interventional C-arm CT images onto live fluoroscopic images. Methods The workflow of the proposed method includes an initialization step, followed by updating overlay of the aortic root model. In the initialization step, a 3-D geometrical mesh model of the aortic root including anatomical valve landmarks is imported from the fluoroscopic and angiographic C-arm CT system [2]. The projected aortic mesh model is then manually aligned onto a contrast image. The contrast image is a fluoroscopic image which is automatically detected and displays well the aortic root roadmap filled with contrast agent. To follow the motion of the aortic root within live fluoroscopic images, the updating of overlaid aortic root model is divided into two image-based tracking procedures as follow. Template-based tracking of a pigtail catheter is applied to approximate translational motion of the aortic root without contrast agent in X-ray image sequences [3]. The collaborating physicians confirmed that the aortic root moves with the pigtail catheter unless it is pulled or pushed manually. For fluoroscopic images with contrast agent, the template-based tracking procedure is temporarily stopped, because the pigtail catheter is almost hidden and cannot be tracked. Rigid intensity-based registration framework of the Insight Segmentation and Registration Toolkit (ITK) has therefore been used to align the predefined contrast image including the initialized position of overlaid aortic root model as a moving image with input contrasted fluoroscopic images as fixed images [4]. The result of 2-D/2-D image registration provides new position values for the overlaid aortic mesh model onto live contrasted X-ray images. Once the contrast agent disappeared in input images of the sequence, the registration-based tracking procedure is deactivated; consequently the template-based tracking procedure works again. Results Retrospective experiments were carried out on several datasets of six females and four males patients under the surgical procedures of trans-apical TAVI. All patient datasets included X-ray image sequences and related aortic root models. Figure 1 (left) depicts manual alignment of projected aortic mesh model with the aortic roadmap onto the contrast image for initializing model-updated overlay algorithm. A visual comparison between static and dynamic overlay of the aortic root model onto one X-ray fluoroscopic image of the same sequence has been shown in Fig. 1 (middle) and Fig. 1 (right) respectively. For X-ray images without contrast agent, there are no tools or ground truth datasets to identify the accurate overly of aortic root models. Therefore, the performance of updated aortic root model overlay onto fluoroscopic images was indirectly evaluated by calculating the absolute mean and maximum displacement errors of the pigtail template over all tested datasets. The maximum and mean
123
Int J CARS (2013) 8 (Suppl 1):S115–S118
Fig. 1 Results of overlaid aortic root mesh model including valve landmarks onto X-ray fluoroscopic images of one patient dataset during the trans-apical TAVI: Initialized overlay onto a contrast image (left), static overlay (middle), and dynamic overlay (right) displacement errors of the pigtail catheter are \1.8 and 1.0 mm respectively. Based on registration-based tracking procedure, the tested image datasets showed dynamic overlay of the aortic root model with relatively small mean displacement errors \1.4 mm and high maximum displacement errors approximately 1.98 mm. Conclusion The results of proposed image-based tracking method demonstrated a successful performance to update the overlaid aortic root model onto live X-ray fluoroscopy for assisting surgical procedures of the TAVI. They do not require any additive optical or electromagnetic tracking systems, which may complicate the surgical workflow. The updated overlay accuracy is expected to be improved by using a 3-D?t aortic root model, showing exact deformations of the aortic root during minimally invasive cardiac procedures. References [1] Walther T, Schuler G, Borger MA, Kempfert J, Seeburger J, Ru¨ckert Y, Ender J, Linke A, Scholz M, Falk V. et al. (2010) Transapical aortic valve implantation in 100 consecutive patients: comparison to propensity-matched conventional aortic valve replacement. European Heart Journal,31(11):1398–1403. [2] Yefeng Z, John M, Rui L, Nottling A, Boese J, Kempfert J, Walther T, Brockmann G, Comaniciu D (2012) Automatic Aorta Segmentation and Valve Landmark Detection in C-Arm CT for Transcatheter Aortic Valve Implantation. Medical Imaging, IEEE Transactions on, 31(12):2307–2321. [3] Karar M, John M, Holzhey D, Falk V, Mohr F–W, Burgert O (2011) Model-Updated Image-Guided Minimally Invasive OffPump Transcatheter Aortic Valve Implantation. In:Medical Image Computing and Computer-Assisted Intervention - MICCAI 2011. Volume 6891, edn. Edited by Fichtinger G, Martel A, Peters T: Springer Berlin Heidelberg; 2011: 275–282. [4] Merk DR, Karar ME, Chalopin C, Holzhey D, Falk V, Mohr FW, Burgert O (2011) Image-guided transapical aortic valve implantation: sensorless tracking of stenotic valve landmarks in live fluoroscopic images. Innovations (Phila),6(4):231–236.
