iros2019国际学术会议论文集pdf0487

1、2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) Macau, China, November 4-8, 2019Robotic Ultrasound for Catheter Navigation in Endovascular ProceduresFernanda Langsch,1, Salvatore Virga,1, Javier Esteban1, Rudiger Gobl1 and Nassir Navab1,2Abstract Endovascular procedur

2、es require real time visual feedback on the location of inserted catheters. This is currently achieved using X-ray fluoroscopy, which causes exposure to radiation. This study describes an alternative method using a robotic ultrasound system for catheter tracking and naviga- tion in endovascular inte

3、rventions, focusing on endovascular aneurysm repair. This approach relies on the registration of pre-operative images to provide both a tracking trajectory and visual feedback of the real-time catheter position. The proce- dure was validated on healthy volunteers and on a phantom that included a rea

4、listic vessel structure, showing an average tracking error of the moving catheter tip of 1.78 1.02 mm.I.Abdominal aortic dition characterizedINTRODUCTIONaneurysm (AAA) is a vascular con- by the enlargement of a portion ofthe aorta, leading to weakening of its walls and possiblevessel rupture. With a

5、 global incidence rate ranging between 4% and 11% per year and a mortality rate of 80% 1, AAA poses a significant health risk for a vast proportion of the population, particularly males aged over 65 years. AAA is typically treated via Endovascular Aneurysm Repair (EVAR), a minimally invasive procedu

6、re in which a catheter is guided to the aneurysm site and is used to deploy a stent graft that captures the blood flow. This reduces the mechanical stress on the vessels walls and prevents rupture 2. Pre-surgical planning relies heavily on pre-operative images, typically computed tomography angiogra

7、phy (CTA), to determine the appropriate dimensions of the stent graft and manufacture it accordingly. During the procedure, the stent graft is guided with precision to a location that guarantees complete coverage of the aneurysm, while occlusion of other vessel branches, including the renal arteries

8、, must be avoided. For EVAR procedures, fluoroscopy imaging is the standard modality for catheter localization and navigation. However, this technique causes exposure to ionizing radiation for both the patient and medical staff and requires the use of a contrast agent. Additionally, pre-operative im

9、aging is typically not integrated into intra-operative navigation, compelling the sur- geon to construct a mental map between the two modalities, fluoroscopy and CTA.In contrast to fluoroscopy, ultrasound (US) is a widely available non-ionizing imaging modality that permits a clear view of the aorta

10、, and in some countries, it is already employed in screening programs for AAA diagnosing 3.Fig. 1.Overview of the system. Top: autonomous roboticUS onvolunteer; bottom left: catheter view in US; bottom right: navigation viewof catheter location. Please refer to the provided supplementary material fo

11、r an additional view of the proposed solution in action.However, US has the limitation of being a highly user- dependent modality, thus yielding poor reproducibility. In contrast, robotic systems permit precise and repeatable ac- quisitions and provide the tracking information needed to obtain 3D US

12、 volumes from conventional 2D scanners.The introduction of mechatronic platforms aimed at au- tomatizing US acquisitions and allowing telemanipulation of US transducers has been the focus of research for almost two decades 4. The research efforts in the field extend from the design of specific end-e

13、ffectors for the steering of US probes 5 to the introduction of complete systems for imaging of arbitrary human anatomies using magnetic resonance imaging (MRI)-based trajectories 6. Specifically for AAA, a robotic solution for autonomous diagnosis using US has been proposed 7. An overview of the st

14、ate-of-the- art in robotic US research is described in 8.When considering US-based guidance for endovascular procedures, the main goal is to provide the surgeon with the real-time position of the inserted catheter based on its detection in the available images. Approaches for visual servoing of medi

15、cal robots based on live tracking of sur- gical devices have been proposed for various scenarios. The automatic alignment of an US probe to the insertion path of a surgical needle was described in 9, while in 10 the tip of a flexible needle was followed by a robotic systemThese authors contributed e

16、qually to this work.1Computer Aided Medical Procedures, Technische Universitat Munchen,Munich, Germany fernanda.langschtum.de2Computer Aided Medical Procedures, Johns Hopkins University, Bal- timore, MD, USA978-1-7281-4003-2/19/$31.00 2019 IEEE5404Fig. 2. Diagram of the system workflow. A 3D US volu

17、me acquired with robotic US and its detected vessel centers (US Vessel Detection), together with the pre-operative image data go through registration (intensity-based and feature-based), yielding the location of vessel centers in the robot frame (Two-step Registration). The robot performs catheter t

18、racking by moving over the computed vessel locations (Robot US Servoing). Image-based catheter detection is facilitated by reducing the search space with knowledge of the vessel location (US Catheter Detection). Finally, the catheter position is displayed over the segmented vessel for navigation (Vi

19、sual Navigation).for minimally invasive procedures. Other approaches, focus on the direct servoing of steerable catheters based on visual information obtained from 3D US 11. A review of visual servoing techniques in medical robotics is also available 12. Most notably, a robotic system for US-based t

20、racking of manually inserted catheters has been presented in 13. This system comprises a robotic manipulator equipped with a linear transducer and a custom catheter with a built-in US active element at its tip. Tracking of the catheter was performed by detecting the US signal received by the single

21、piezoelectric element mounted on the catheter rather than by visual processing. While this technique simplifies the tracking problem and may outperform vision-based tracking, it requires modifications to standard clinical catheters and in- tegration of additional hardware. Additionally, the proposed

22、 method, as presented in 13, does not allow for a seamless bi-directional tracking of the catheter position and does not combine the tracking information into any form of visual navigation for the user.and contours of a human body. It was shown that the system was able to follow the bi-directional c

23、atheter motion and visualize its current position in a precise manner.II. METHODSThe following section describes the individual components that enable the system to perform catheter navigation for EVAR procedures using a robotic US platform. The re- quired preprocessing of the pre-operative data is

24、described in Sec. II-A, and Sec. II-B provides details on matching of the computed path along the vessel centerlines to the current patient position. Sec. II-C presents the image-based catheter detection employed and the corresponding visual servoing strategy. An overview of the system components an

25、d their interaction is shown in Figs. 1 and 2.A. Data PreprocessingThe abdominal vessel tree was segmented from an input pre-operative volume, and its centerlines were computed. This includes the abdominal tract of the aorta, a short portion of the iliac arteries after the aorta bifurcation, and the

26、 main renal arteries. Segmentation and centerline extraction were performed using the Vascular Modelling Tool Kit (VMTK)1, based on active contour level set segmentation and on a weighted geodesic search over a Voronoi diagram 14, respectively. The output of this process was a set ofcenterline point

27、s, C R3, which was then used to guide the robotic motion during the catheter tracking phase. That is,the catheter position was naturally constrained by the vessel geometry, thus, the robot motions and catheter tracking was limited to the computed locations.B. Patient Registration1) US Compounding: A

28、 regular-spaced 3D US volume was obtained by combining the real-time tracking infor- mation of the US transducer supplied by the robot and the 2D US images. A backward-normalized convolutionIn this study, we present a complete robotic US sys- tem targeted at integrating US-based catheter tracking an

29、d visualization to enable radiation-free navigation for EVAR procedures. To achieve this, standard pre-operative volumes were processed to segment the vessel structure of interest and obtain its centerlines. 3D US volumes of the target anatomy were registered to the pre-operative data to create a co

30、rrespondence of the known vasculature to the intra- operative setting. The computed vessel path was then used to guide the US probe during the manual insertion of a catheter. Image-based catheter detection within the US frames is employed for real-time tracking, allowing for retrieval of the cathete

31、r position along its known path. Finally, the location of the catheter tip was visualized within the segmented vessel structure, providing intuitive navigation feedback to the end user. The proposed registration scheme was validated on simulated catheter motions for four healthy volunteers, and the

32、complete system was demonstrated on a custom phantom that realistically mimicked both the tissue structure15405Fig. 3. Vessel detection steps. a) Original US image, b) vesselness response using the modified Frangi filter, c) intensity dip mask, d) binary mask applied to the vesseln

33、ess response and e) estimated vessel centerline by ellipse fitting.technique was employed, as described in 15. The initial intra-operative acquisition can be performed by moving the robotic arm equipped with a US probe manually or by executing a planned trajectory autonomously as performed in 7.(0,

34、if 2 0otherwiseV (s) =(1)R2S2exp(1 exp ( 2c2 )B22withq|1|222) Image-based Registration: To spatially align the two available volumes (pre- and intra-operative) an image-based registration based on the LC2 similarity measure 16 is used. This is beneficial to cope with the two pre-operative modalities

35、 used for the proposed validation, i.e., CT for a phantom and MRI for human volunteers. In fact, in the US/CT case, US intensities are correlated to the output of an US simulation from the CT volume, whereas in the US/MRI case, MRI intensities and gradient magnitudes are associated with US intensity

36、 values. In contrast to 16, we performed a rigid registration alone rather than a deformable registration. A deformable registration using the same LC2 similarity method is feasible, however, it remains difficult due to high soft tissue deformation in human US acquisitions. LC2- based registration w

37、as employed to provide an initialization for the feature-based method described in II-B.4, which focuses on the optimal alignment of the vessel structures visible in both modalities.RB =and S =1+ ,2(2)| |2where |1| |2| are the eigenvalues of the Hessian matrix of the US image and and c are constants

38、.Since our 2D images provided cross-sectional views of the vessels, the vesselness response required altering so that it provided strong responses to circular structures instead oftubular ones. A modification to the definition of RB is hereby proposed asRB = |2 1| .(3)That is, the original formulati

39、on of Rwould lead toBa high response for elliptical structures (i.e., |1| |2|), whereas our variation leads to a high response so for circularstructures (i.e., 1 2). Additionally, enforcing V (s) = 0, if 1, 2 ,(10)with k = 150 and = 5 106 in our case. Accordingto the detection response, the index of

40、 the centerline point guiding the robot motion was updated as follows:(eit+ 1 if match in Iit 1 else .i=(11)t+1An additional indexing was used for the tracking of the catheter tip location,(iif match in Ietjt+1 =(12)it 1 else .That is, while the robot will continuously be in motion to capture any ca

41、theter movements in both possible directions, the location of the catheter is updated to the centerline point, cj, which is closest to the location of its last detection. As a remark, the accuracy of the catheter location can be increasedd = ci+1 ci .(4)Fixing the Z-axis toTz = (0, 0, 1) ,(5)the X-a

42、xis was obtained as the vector rejection of d on z(dx,dy, 0)T R3 ,x =(6)|(dx, dy, 0)T |whereas the Y-axis was simplyy = z x .The final transducer pose was composed as(7)Fig. 4. Catheter detection. Synthetic template (left), selected ROI (center) and response from template matching (right).P = xyzci

43、.(8)5407Fig. 5. Experiments are performed on a phantom mimicking a human figure (left) and containing a realistic vessel tree (right). Vessel branches marked in green represent the femoral arteries, in blue the abdominal aortic tract and in red the renal arteries.Fig. 6. Tracking validation. The cat

44、heter tip location computed by the system (yellow dot) is compared to a manual annotated tip location (red line). Distances are computed along the centerline.in Sec. II-A (Data Preprocessing) and Sec. II-B (Patient Reg- istration) was performed in all phases to achieve an optimal registration between the vessel centerlines extracted from the MRI and US volumes. Next, autonomous robotic US servoing was performed using simulated cath