A pre-operative planning framework for global registration of laparoscopic ultrasound to CT images

A pre-operative planning framework for global registration of laparoscopic ultrasound to CT images Purpose Laparoscopic ultrasound (LUS) enhances the safety of laparoscopic liver resection by enabling real-time imaging of internal structures such as vessels. However, LUS probes can be difficult to use, and many tumours are iso-echoic and hence are not visible. Registration of LUS to a pre-operative CT or MR scan has been proposed as a method of image guidance. However, the field of view of the probe is very small compared to the whole liver, making the registration task challenging and dependent on a very accurate initialisation. Methods We propose the use of a subject-specific planning framework that provides information on which anatomical liver regions it is possible to acquire vascular data that is unique enough for a globally optimal initial registration. Vessel-based rigid registration on different areas of the pre-operative CT vascular tree is used in order to evaluate predicted accuracy and reliability. Results The planning framework is tested on one porcine subject where we have taken 5 independent sweeps of LUS data from different sections of the liver. Target registration error of vessel branching points was used to measure accuracy. Global registration based on vessel centrelines is applied to the 5 datasets. In 3 out of 5 cases registration is successful and in agreement with the planning. Further tests with a CT scan under abdominal insufflation show that the framework can provide valuable information in all of the 5 cases. Conclusions We have introduced a planning framework that can guide the surgeon on how much LUS data to collect in order to provide a reliable globally unique registration without the need for an initial manual alignment. This could potentially improve the usability of these methods in clinic. Keywords Laparoscopy · Laparoscopic ultrasound · Surgical planning · Rigid registration · Feature-based registration · Global registration Introduction cost savings for the healthcare system due to shorter hospital stays [2]. However, only 5–30% of patients are considered Liver cancer is a major health problem and 150,000 patients for this approach. In the case where the target tumours are per year could benefit from liver resection [1]. Laparoscopic too large or close to critical vascular regions, the procedure liver resection provides benefits over open surgery such as is considered high risk given the limited field of view and reduced pain and faster recovery for the patient, along with lack of haptic feedback [3]. Laparoscopic ultrasound (LUS) can potentially decrease this risk since it images sub-surface structures such as vessels and some tumours. However, many B João Ramalhinho tumours are iso-echoic and hence are not visible in the joao.ramalhinho.15@ucl.ac.uk ultrasound images. Therefore, registration of LUS to a pre- Wellcome/EPSRC Centre for Interventional and Surgical operative CT scan using the signal from vascular structures Sciences, University College London, London, UK has been proposed as an image-guidance method. Centre For Medical Image Computing, University College Registration of LUS to CT is a very challenging task London, London, UK because in addition to the liver being deformed during LUS Division of Surgery and Interventional Science, University imaging, the imaging field of view is small and restricted College London, London, UK 123 1178 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 by the limited freedom of movement of the LUS probe. The To overcome this operator dependence on initialising the problem then becomes the alignment of a partial subset of registration, a possible solution is to perform a globally the liver to a whole CT scan volume. If the subset is too small optimised initial alignment. Given the topology of the liver and contains for example an individual vessel segment, the vascular tree, we hypothesise that it is possible to globally registration problem is poorly constrained and depends on a align a small LUS volume to the large CT volume if the vas- very accurate initialisation. However, our hypothesis is that culature captured in the LUS is geometrically unique. The given enough ultrasound data, a set of vessels and their rel- remaining challenge then becomes knowing in which areas ative pose and shape is likely to have a unique solution. The of the liver we can find such subsets. Therefore, we propose task of this paper is to determine what volume needs to be a pre-operative planning framework to tell the surgeon both scanned in order to obtain a unique vascular configuration as where these regions are and how much vascular data in their a function of location across the liver. This information can vicinity is required for a unique alignment. then be used to guide the surgeon to the appropriate liver vol- ume that needs to be acquired in order to achieve a reliable Contributions of this paper registration. To the best of our knowledge, there is currently no method to Background plan LUS acquisition in terms of registration outcome. Our contributions include: The Ultrasound (US) to CT registration problem has been tackled by several groups. Solutions for freehand US reg- – A comparison study on the assessment of which liver vascular features are best for registration. istration have been proposed either by matching US to a combination of the CT signal and a simulated US signal [4], – A framework that predicts how much data must be col- or by registering vessel probability maps from both US and lected at each region of the liver for a globally optimal CT [5]. These methods have been demonstrated using images registration. that capture large sections of liver and our preliminary tests – A proposed global registration approach for LUS to CT suggest such intensity based methods do not work well on the data, with results that are compared to the planning. restricted volume captured by the LUS probe. Other authors aligned 3D US instead, using reconstructed vessels as fea- tures [6,7], vessels and liver surface [8,9] or image intensity Vascular features assessment gradient information [10]. However, there are currently no 3D probes for laparoscopy on the market. In order to understand the nature of the vascular data that is to Few authors have investigated the specific registration of be used in our planning framework, we perform a preliminary LUS to CT. The first feasibility studies were proposed by Bao study that compares the registration outcome when using two et al. in an isolated phantom [11] and Kruecker et al. [12]on distinct vascular features: the vessel branching points, also a complete experimental setup. The first animal results in know as bifurcations, and the vessel centreline points. an ex-vivo porcine liver were reported by Martens et al. [13] For this purpose, we simulate the perfect acquisition of using a surface-based rigid registration with the LapAssistent subsets of N bifurcations and surrounding neighbouring system. During surgery, besides being moved by breathing centrelines, apply displacements to simulate expected defor- motion, the liver is compressed by the LUS probe upon mations or segmentation inaccuracies, and register them back contact and deformed by pneumoperitoneum (abdominal to the total model (see Fig. 1). Bifurcations are registered insufflation required for laparoscopy). This makes globally using point-based registration [17], mimicking the scenario rigid solutions such as the three aforementioned ones not where a surgeon picks common landmarks in both modalities sufficient. Even though deformable solutions are available and performs a rigid alignment. Centrelines are registered [7,8,10,14,15], they are hard to validate clinically and may using Iterative-Closest-Point (ICP) [18,19], simulating the require unfeasible computation times during surgery. Song case where correspondences are not known, and the best et al. [1] proposed locally rigid registration to be sufficient rigid solution using segmented centrelines is computed. In given a small enough liver region of interest, and validated the case where there are no displacements, the resulting trans- a vessel-based approach on in-vivo porcine data. However, formation is the identity. Otherwise, a new transformation is the method required manual selection and matching of vessel obtained and applied to the whole liver model. Using the bifurcations between LUS and CT, a task that is very diffi- new bifurcation positions and the original ones, we compute cult and time-consuming during surgery. Previously, in [16], a simulated Target Registration Error (TRE) over all bifur- we attempted vessel-based registration without establishing cations and assess the performance of both approaches. landmark correspondences, but the results still depended on We use this method to test the influence of the number of the accuracy of the initialisation. picked bifurcations and amplitude of applied displacements. 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1179 Fig. 1 Assessment of registration based on the two different vascular right are the resulting configuration, green the original one. Colormaps features. Point-based registration on bifurcations and ICP on centrelines show predicted displacements between these two, values ranging from are used to recover a transformation T . The red vessel models on the blue (low) to red (high) 50 combinations of N landmarks are randomly selected for N surrounding 3 bifurcations, and without specifying point cor- ranging from 3 to 10. For each combination, registrations are respondences. This indicates that within a liver region, fewer computed after applying a set of 1000 Gaussian distributed vascular branchings are required by the centreline ICP to displacements to the selected features, 200 for each devia- yield an accurate result. Since in a real scenario LUS reg- tion σ ranging from 1 to 5 mm. In the bifurcation approach, istration to CT needs to be accurate in a local region of displacements are applied in each point separately. For the interest, we also assess the accuracies in terms of a confi- centrelines, to better simulate deformation between branches, dence margin. For each registration, we compute the centre a random displacement vector is applied to all the points that of the picked points and find the maximum distance from it compose a branch. below which we capture bifurcations with TRE below 5 mm. The results of this experiment are displayed in Fig. 2. Each In the bottom row of Fig. 2, we present the mean results of curve describes the TRE results for a fixed number of bifur- this radial distance the same way as in the accuracy plots. In cations given different amplitudes of displacements. In the both methods, an increasing number of bifurcations increases top row, the Root Mean Square (RMS) of the bifurcation this distance. However, unless deviations surpass 3 mm, the TRE over all the vascular tree is presented in millimetres. confidence margins of ICP are superior to point-based reg- Each point in a curve averages a result of 200 registra- istration: ICP with centrelines around only 3 bifurcations tions over 50 bifurcation configurations. As expected, in both gives a wider mean working radius than point-based regis- methods the more bifurcations used, the lower the resulting tration on 5 bifurcations. Picking the previous example of 6 error. Considering 5 mm to be a clinically relevant level of landmarks and 1 mm of deviation, there is an improvement of a accuracy, we can see that using point-based registration 110–150 mm of the working radius. Therefore, besides being would require manual identification of 6 bifurcations with more accurate, centreline ICP also requires fewer branching an accuracy of 1 mm. This is observed in the green curve points to be reliable. with displacement σ = 1 mm. However, to obtain the same These results suggest that up to 6 landmarks would be result, the ICP approach would only require the centrelines required for a point-based registration. Clinically, identify- 123 1180 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 Fig. 2 Accuracy results on vessel bifurcations after running results of than 5 mm. The left column refers to bifurcation based registration; the the first experiment. The top row presents the RMS of the TRE on bifur- right column refers to centreline-based registration cations; the bottom row represents the radius where accuracy is better ing this number of landmarks during surgery would be very Uniqueness assessment challenging and time-consuming for a surgeon, making the vessel centreline approach preferable. We therefore conclude Our uniqueness assessment strategy relies on the use of any that vessel centrelines will provide a more clinically relevant Global ICP method, such as the Globally Optimal ICP (GO- registration result than bifurcations. ICP) proposed by Yang et al. [20]. This algorithm searches the solution space of rotations and translations with a Branch and Bound approach to find the rigid alignment that best minimises the L norm between two point clouds given a Methods minimum error threshold. We hypothesise that a subset of vessel centreline points is unique if GO-ICP aligns them cor- Given the demonstration above that centrelines should pro- rectly with the original complete tree using a sufficiently vide better registration results than bifurcations, our frame- small error threshold. Since the problem is rigid, the global work aims to assess in a planning stage which centreline point minimum of the distance function is 0, in which the subset sets are unique and possible to be aligned globally without is perfectly aligned in the original position. However, by set- initialisation. This assessment is performed over the liver sur- ting the threshold limit to a higher value, we allow GO-ICP to face, guiding a surgeon to acquire the necessary amount of find solutions with larger distance error. When the algorithm ultrasound data to acquire a unique set of vessels. 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1181 Fig. 3 Framework for assessment of uniqueness radius in a surface point of the pre-operative CT. Green represents points captured by a sphere around a point sampled on the pre-operative surface. Red represents GO-ICP resulting alignment reaches this value and outputs a configuration that is not close delineated in order to obtain binary masks of vessel sections; to the original one, the distance function has at least one local secondly, centreline points are estimated as the centroids of minimum that is not related to the global one, implying that each of these vessel sections. This segmentation strategy aims the subset is not geometrically unique. to simulate results that could be obtained with an automatic framework using only intensity information from 2D images. Currently, such approaches could be implemented for exam- Uniqueness planning ple using Deep Learning methods. An intra-operative probe contact position P of the dataset is computed as the mean Instead of using complete centreline branches as in “Vascular of the tracked surface digitised points as defined in [16]. features assessment” section, our framework evaluates the The same expanding radius strategy of “Uniqueness plan- uniqueness of centreline point subsets captured by simulated ning” section is then applied to the segmented centrelines probe positions along the pre-operative model surface. This using P as the sphere centre. In each registration, the TRE process is illustrated in Fig. 3. Firstly, the surface subset that is anterior and visible during surgery is extracted manually between manually picked bifurcations from both the LUS dataset and CT is measured. We define the success radius R and subsampled to generate possible probe contact positions. as the radius from which the TRE result obtained by GO- A sequence of spheres with increasing radius is generated ICP reaches a minimum and stabilises. If this minimum is around each of these positions, and the centreline points that not reached, we consider the registration to be unsuccess- are captured in its domain are tested for uniqueness with ful. GO-ICP. We then define the uniqueness radius R as the In the real LUS data several deformation components minimum value at which the captured centreline points are are expected. This means that the minimum L error found unique. By repeating this process in all the sampled points, a by GO-ICP may not be physically correct. To mitigate this R map is generated which tells the surgeon how broad the effect, we use some prior knowledge of the LUS acquisition acquisition must be in each liver region for a globally unique to constrain GO-ICP solution search space. For translations, registration. we constrain the dataset to the liver lobe where it was cap- tured, left or right. For rotations, we define a rough estimate LUStoCTregistration of the direction in which the sweep was taken in the CT, and ◦ ◦ allow only for rotations between [− 60 , 60 ]. This strategy To validate the uniqueness planning, we register continu- prevents the algorithm from considering physically impossi- ous sets of 2D tracked LUS images to the pre-operative CT ble acquisition configurations. using the same GO-ICP methodology of the previous sec- To compare the result with the planning, we obtain a tion. For each set, vessel centrelines are extracted from the reference standard transformation by applying point-based images in two steps: firstly, vascular structures are manually 123 1182 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 Fig. 4 Uniqueness map obtained by sampling 400 points over the visible surface. Sphere radius ranges from 5 to 100 mm in steps of 5 mm. Red and blue represent high and low values of uniqueness radius respectively registration in the landmarks used for TRE assessment. The face results to the whole visible surface points to obtain a aim here is to test how well our uniqueness map predicts complete uniqueness radius map. the amount of acquired data required for a reliable registra- Global registration experiments are performed in 5 sets of tion. Therefore, the radius R at which GO-ICP successfully LUS images, 3 acquired from the right lobe of the subject aligns the LUS data to the CT should be similar to the radius and 2 acquired from the left. In order to reduce fluctua- R predicted by the framework given the point-based refer- tion effects from the EM tracking system, we fit a cubic ence standard alignment. spline to the digitised surface and apply the result to the centreline points, smoothing them. We run the uniqueness assessment with the same parameters as the planning plus Experiments the prior solution space constraints described in “LUS to CT registration” section. Reference standard point-based align- Experiments were performed in retrospective data from a ments are obtained separately for each dataset. By applying single porcine subject previously acquired in [1]. Surface and the resulting transformation to the respective positions P, vascular models were segmented from standard abdominal reference probe positions P are obtained and used for com- triphase CT scans with resolution 512 × 512 × 2.5 mm. LUS parison with the generated R map. images of resolution 384 × 456 mm were acquired at a rate 2 3 of 10Hz using an Analogic SonixMDP and a Vermon LP7 linear probe which was electromagnetically (EM) tracked by Results an NDI Aurora tracking system at a rate of 40Hz. We test the planning framework by sampling 400 probe Uniqueness planning contact positions evenly spaced along the surface using the Farthest-point sampling algorithm [21]. We vary the sphere Figure 4 shows the obtained uniqueness radius map. Red radius in steps of 5 mm in the range [5–100] mm. We set the areas have a high uniqueness radius, implying that broader minimum threshold of GO-ICP to a mean distance error of data collections around them are required for a unique reg- 2.4 mm. Smaller values give finer results but require larger istration. In contrast, blue regions have a lower uniqueness computation times. We linearly interpolate the sampled sur- radius and require less data. The highest values are obtained at the liver edges, which is expected given that vasculature is less present in the liver periphery. The best and lowest val- www.visiblepatient.com. ues are observed in more central regions of the liver, ranging www.analogicultrasound.com. from 25 to 40 mm. The exception is the liver top surface www.vermon.com. which is more distant from the vasculature, yielding results www.ndigital.com. around 50 mm. 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1183 Fig. 5 Accuracy results of LUS to CT global registration experiments. Left represents results using a normal CT scan as target; Right represents results using CT scan under insufflation. Yellow markers represent the stable minimum that is measured as R LUStoCTregistration registration experiments using an insufflated CT vascular tree and therefore compensate for pneumoperitoneum. CT scans LUS to CT registration accuracy results are displayed in under these conditions are not acquired clinically, and hence the left plot of Fig. 5. For each dataset, the RMS of TRE are not used for planning. However, we present these results is presented as a function of the capture radius around the for illustrative purposes. Accuracy results of this experiment respective intra-operative probe contact position. For the are presented in the right-hand side of Fig 5. along with cases where TRE stabilises at a minimum and registration the respective radius measurements in the right-hand side of is successful, a yellow marker highlights the corresponding Table 1. In this case, all datasets are registered successfully success radius R . Table 1 summarises the obtained success with a R radius similar to the planning, with differences S S radii R along with the corresponding R predicted by the ranging from 0.5 to 9.1 mm. Furthermore, in agreement with S U planning framework. For each dataset, this R value is taken the successful cases of the normal experiments, the distances as the uniqueness radius measured at the planning surface D never surpass 16 mm. point closest to the reference probe position P . Error E Visually, we can observe this improvement in Fig. 6.In R S represents the RMS of TRE measured at radius R .Error E each row, the result of each of the three registration exper- S R represents the RMS of TRE obtained by the reference stan- iments is presented. We can observe that for Dataset 5, dard. Since this standard is the result of direct minimisation GO-ICP with the measured R of 30 mm aligns the recon- of the distance between the landmarks with a rigid solution, structed ultrasound sweep in the same region as the reference E provides the minimum TRE error that can be achieved. standard. The same happens if the insufflated scan is used, It is possible to see that 3 out of 5 cases were successfully with the R of 40 mm. In the case of Dataset 4, we observe registered, and that the respective R values are similar to that registration to the normal scan is not successful with the prediction R , with differences of 1.5, 7.2 and 5 mm. In any radius, but with an insufflated scan results improve and the two failure cases, registration is never successful, and no a radius R of 40 mm is observed. The same improvement comparison with planning can be made. To better understand from normal to insufflated scan was obtained with Dataset 3. these cases, we measure D , the distance between P and R R corresponding R planned location. In the successful cases, this distance does not surpass 15 mm. However, in the failure Discussion cases, this result surpasses 23 mm, indicating a large differ- ence between intra and pre-operative surface and therefore The results of our uniqueness mapping are as expected, since implying the presence of significant liver tissue deformation. surface regions that are far from the vasculature are predicted In order to assess how much deformation is influencing to require broader data acquisitions. In practical terms, the this registration problem, we repeat the same planning and mapping of Fig. 4 would guide a surgeon to acquire data 123 1184 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 Table 1 Success radius Dataset Normal CT scan CT scan under insufflation measurement results of LUS R R E E D R R E E D U S S R R U S S R R datasets 1 (Left lobe) 31.5 30.0 16.3 5.94.8 39.1 30.0 20.98.613.2 2 (Left lobe) 32.2 25.0 10.4 1.814.8 40.0 35.0 21.82.82.4 3 (Right lobe) 45.0 – – 4.725.0 43.8 40.0 9.66.815.8 4 (Right lobe) 40.0 – – 21.323.1 40.5 40.0 15.310.314.4 5 (Right lobe) 35.0 30.0 12.8 9.413.2 45.0 40.0 5.23.29.1 E represents the RMS of TRE obtained with registration at success radius R ; E represents the same error S S R for the reference standard transformation; R represents the uniqueness radius predicted by the framework using the reference standard; D represents distance between the reference probe position P and surface R R location of R measurement. All values are in millimetres Fig. 6 Visual results of registrations of LUS datasets 4 and 5 from the GO-ICP registration failure using maximum radius; Right shows GO- right lobe. Left shows the reference standard results. Middle top shows ICP result with insufflated scan as target at the measured success radius GO-ICP result with measured success radius; Middle bottom shows preferentially in the blue areas, where smaller volumes of cies ranging from 5.2 to 21.8 mm. Higher errors above 20 mm data are required for a globally optimal alignment. and much larger than the reference standard are observed in In the LUS to CT registration experiments, 3 out of the left lobe datasets. This may be explained by errors in 5 datasets were registered successfully with accuracy val- the manual localisation of bifurcations. Furthermore, retro- ues E of 16.3, 10.4 and 12.8 mm (Table 1 column 3). spective analysis showed that the CT insufflated scan was In the two failure cases, large differences between intra- unreliably segmented in this region. Regardless, the datasets operative and pre-operative surface positions indicate defor- are aligned in the correct region, but with a larger translation mation. From the three expected deformation factors listed error. These results indicate that this registration problem in “Background” section, surface compression and pneu- may be more affected by insufflation than surface compres- moperitoneum are the ones that may better explain this sion. phenomenon. Further experiments with an insufflated target In the normal scan experiments, for the 3 successful cases, led to successful registrations with all datasets with accura- the planned radius R is similar to the measured success 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1185 radius R with a maximum difference of 7.2 mm. In the Acknowledgements JR, MRR were supported by the EPSRC CDT in Medical Imaging [EP/L016478/1]. MJC, DH, DB, ST were supported insufflated case, a larger difference of 9.1 mm is observed for by the Wellcome/EPSRC [203145Z/16/Z]. BD, DH was supported by dataset 1, a fact which is possibly related to the lower accu- the NIHR Biomedical Research Centre at University College London racy of the registration. However, it can be observed that Hospitals NHS Foundations Trust and University College London. The the radii R are always lower than the planned R . Since imaging data used for this work was obtained with funding from the S U Health Innovation Challenge Fund (HICF-T4-317), a parallel funding we corrected for insufflation in the second experiment, this partnership between the Wellcome Trust and the Department of Health. trend could be related to surface compression—if the surface The views expressed in this publication are those of the author(s) and not is pushed closer to the vessels, unique vasculature would be necessarily those of the Wellcome Trust or the Department of Health. found with a lower capture radius. Regardless, the difference between predictions and measurement is not critical as a sur- Compliance with ethical standards geon would not be able to collect data with a 5 mm radius precision. In addition to showing a reasonable agreement with the Conflict of interest The authors declare that they have no conflict of planning predictions and the obtained results, this approach interest. points to the feasibility of GO-ICP method with prior knowl- Ethical approval All applicable international, national, and/or institu- edge for vessel-based registration of LUS to CT. The obtained tional guidelines for the care and use of animals were followed. All TRE values obtained in the normal experiment are promis- procedures performed in studies involving animals were in accordance ing given the fact that we aimed to initially align rigidly a set with the ethical standards of the institution or practice at which the studies were conducted. For this type of study, formal consent is not of vessels in its correct region. Obtaining clinically valuable required. This article does not contain patient data. accuracies below 10 mm could be achieved by compensat- ing for deformation locally afterwards. Using non-optimised Open Access This article is distributed under the terms of the Creative code and hardware, the uniqueness map compilation and Commons Attribution 4.0 International License (http://creativecomm ons.org/licenses/by/4.0/), which permits unrestricted use, distribution, clinical registration with GO-ICP took 20 h and 10 min and reproduction in any medium, provided you give appropriate credit respectively. Since the former is a pre-operative off-line step, to the original author(s) and the source, provide a link to the Creative this high value would not be critical to clinical application. Commons license, and indicate if changes were made. Conclusions References We have developed a pre-operative planning framework to 1. Song Y, Totz J, Thompson S, Johnsen S, Barratt D, Schneider guide the surgeon in acquiring LUS data in a way that a C, Gurusamy K, Davidson B, Ourselin S, Hawkes DJ, Clarkson reliable registration to CT can be obtained without an initial MJ (2015) Locally rigid, vessel-based registration for laparoscopic manual alignment. We consider this framework to be relevant liver surgery. In: IJCARS, pp 1–11 2. Ciria R, Cherqui D, Geller DA, Briceno J, Wakabayashi G (2016) since LUS is particularly difficult to handle, and providing Comparative short-term benefits of laparoscopic liver resection: guidance beforehand could increase the use of this regis- 9000 cases and climbing. Ann Surg 263(4):761–777 tration method and enhance the safety of laparoscopic liver 3. Wakabayashi G, Cherqui D, Geller DA, Buell JF, Kaneko H, Han resection. HS, Asbun H, O’Rourke N, Tanabe M, Koffron AJ (2015) Rec- ommendations for laparoscopic liver resection: a report from the The major limitation of this work is the restriction of the second international consensus conference held in morioka. Ann study to one subject. To mitigate this risk, we show variabil- Surg 261(4):619–629 ity by applying our method on data from several different 4. Wein W, Brunke S, Khamene A, Callstrom MR, Navab N (2008) regions of the liver. Another limitation maybe the fact that Automatic CT-ultrasound registration for diagnostic imaging and image-guided intervention. Med Image Anal 12(5):577–585 we do not compensate for any deformation in the framework, 5. Penney GP, Blackall JM, Hamady MS, Sabharwal T, Adam A, leading to registration failures. However, for the purpose of Hawkes DJ (2004) Registration of freehand 3D ultrasound and a globally optimal initial alignment, it is difficult to intro- magnetic resonance liver images. Med Image Anal 8(1):81–91 duce deformation parameters. Since successful registrations 6. Aylward SR, Jomier J, Guyon JP, Weeks S (2002) Intra-operative 3D ultrasound augmentation. In: ISBI, pp 421–424 were achieved with the insufflated scan, rigid may be suffi- 7. Lange T, Eulenstein S, Hünerbein M, Schlag PM (2003) Vessel- cient if insufflation is modelled in the pre-operative CT, as in based non-rigid registration of MR/CT and 3D ultrasound for [22]. navigation in liver surgery. Comput. Aided Surg. 8(5):228–240 Our future work directions are twofold. From a techni- 8. Lee Duhgoon, Nam Woo Hyun, Lee Jae Young, Ra Jong Beom (2011) Non-rigid registration between 3d ultrasound and ct images cal point of view, we intend to simulate insufflation in our of the liver based on intensity and gradient information. Phys Med framework and in a later stage probe compression. In terms Biol 56(1):117 of validation, we aim to apply this framework in human data 9. Nam WH, Kang DG, Lee D, Lee JY, Ra JB (2012) Automatic reg- in order to obtain clinically translatable results. istration between 3d intra-operative ultrasound and pre-operative 123 1186 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 ct images of the liver based on robust edge matching. Phys Med 17. Arun KS, Huang TS, Blostein SD (1987) Least-squares fitting of Biol 57(1):69 two 3-D point sets. IEEE Trans Pattern Anal Mach Intell 5:698–700 10. Lange T, Papenberg N, Heldmann S, Modersitzki J, Fischer B, 18. Besl PJ, McKay ND (1992) Method for registration of 3D shapes. Lamecker H, Schlag PM (2009) 3D ultrasound-CT registration of In: Proceedings of SPIE 1611, Sensor fusion IV: control paradigms the liver using combined landmark-intensity information. IJCARS and data structures, pp 586–606 4(1):79–88 19. Chen Y, Medioni G (1991) Object modeling by registration of mul- 11. Bao P, Warmath J, Galloway R, Herline A (2005) Ultrasound-to- tiple range images. In: IEEE international conference on robotics computer-tomography registration for image-guided laparoscopic and automation, pp 2724–2729 liver surgery. Surg Endosc 19(3):424–429 20. Yang J, Li H, Campbell D, Jia Y (2015) GO-ICP: A globally optimal 12. Kruecker J, Viswanathan A, Borgert J, Glossop N, Yang Y, Wood solution to 3D ICP point-set registration. IEEE Trans Pattern Anal BJ (2005) An electro-magnetically tracked laparoscopic ultra- Mach Intell 38(1):12 sound for multi-modality minimally invasive surgery. Int Congr 21. Peyré G (2011) The numerical tours of signal processing. Adv Ser CARS 1281:746–751 Comput Signal Image Process Comput Sci Eng 13(4):94–97 13. Martens V, Besirevic A, Shahin O, Kleemann M (2010) Lapas- 22. Johnsen SF, Thompson S, Clarkson MJ, Modat M, Song Y, Totz J, sistent: computer assisted laparoscopic liver surgery. In: IFMBE, Gurusamy K, Davidson B, Taylor ZA, Hawkes DJ (2015) Database- vol 22 based estimation of liver deformation under pneumoperitoneum for 14. Dagon B, Baur C, Bettschart V (2008) Real-time update of 3D surgical image-guidance and simulation. In: MICCAI. Springer, pp deformable models for computer aided liver surgery. In: ICPR, pp 450–458 1–4 15. King AP, Blackall JM, Penney GP, Hawkes DJ (2001) Tracking liver motion using 3D ultrasound and a surface based statistical shape model. In: MMBIA 2001, pp 145–152 16. Ramalhinho J, Robu M, Thompson S, Edwards P, Schneider C, Gurusamy K, Hawkes DJ, Davidson B, Barratt D, Clarkson MJ (2017) Breathing motion compensated registration of laparoscopic liver ultrasound to CT. In: Proceedings of SPIE medical imaging http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png International Journal of Computer Assisted Radiology and Surgery Springer Journals

A pre-operative planning framework for global registration of laparoscopic ultrasound to CT images

Free
10 pages
Loading next page...
 
/lp/springer_journal/a-pre-operative-planning-framework-for-global-registration-of-vuBYpS4aC7
Publisher
Springer International Publishing
Copyright
Copyright © 2018 by The Author(s)
Subject
Medicine & Public Health; Imaging / Radiology; Surgery; Health Informatics; Computer Imaging, Vision, Pattern Recognition and Graphics; Computer Science, general
ISSN
1861-6410
eISSN
1861-6429
D.O.I.
10.1007/s11548-018-1799-2
Publisher site
See Article on Publisher Site

Abstract

Purpose Laparoscopic ultrasound (LUS) enhances the safety of laparoscopic liver resection by enabling real-time imaging of internal structures such as vessels. However, LUS probes can be difficult to use, and many tumours are iso-echoic and hence are not visible. Registration of LUS to a pre-operative CT or MR scan has been proposed as a method of image guidance. However, the field of view of the probe is very small compared to the whole liver, making the registration task challenging and dependent on a very accurate initialisation. Methods We propose the use of a subject-specific planning framework that provides information on which anatomical liver regions it is possible to acquire vascular data that is unique enough for a globally optimal initial registration. Vessel-based rigid registration on different areas of the pre-operative CT vascular tree is used in order to evaluate predicted accuracy and reliability. Results The planning framework is tested on one porcine subject where we have taken 5 independent sweeps of LUS data from different sections of the liver. Target registration error of vessel branching points was used to measure accuracy. Global registration based on vessel centrelines is applied to the 5 datasets. In 3 out of 5 cases registration is successful and in agreement with the planning. Further tests with a CT scan under abdominal insufflation show that the framework can provide valuable information in all of the 5 cases. Conclusions We have introduced a planning framework that can guide the surgeon on how much LUS data to collect in order to provide a reliable globally unique registration without the need for an initial manual alignment. This could potentially improve the usability of these methods in clinic. Keywords Laparoscopy · Laparoscopic ultrasound · Surgical planning · Rigid registration · Feature-based registration · Global registration Introduction cost savings for the healthcare system due to shorter hospital stays [2]. However, only 5–30% of patients are considered Liver cancer is a major health problem and 150,000 patients for this approach. In the case where the target tumours are per year could benefit from liver resection [1]. Laparoscopic too large or close to critical vascular regions, the procedure liver resection provides benefits over open surgery such as is considered high risk given the limited field of view and reduced pain and faster recovery for the patient, along with lack of haptic feedback [3]. Laparoscopic ultrasound (LUS) can potentially decrease this risk since it images sub-surface structures such as vessels and some tumours. However, many B João Ramalhinho tumours are iso-echoic and hence are not visible in the joao.ramalhinho.15@ucl.ac.uk ultrasound images. Therefore, registration of LUS to a pre- Wellcome/EPSRC Centre for Interventional and Surgical operative CT scan using the signal from vascular structures Sciences, University College London, London, UK has been proposed as an image-guidance method. Centre For Medical Image Computing, University College Registration of LUS to CT is a very challenging task London, London, UK because in addition to the liver being deformed during LUS Division of Surgery and Interventional Science, University imaging, the imaging field of view is small and restricted College London, London, UK 123 1178 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 by the limited freedom of movement of the LUS probe. The To overcome this operator dependence on initialising the problem then becomes the alignment of a partial subset of registration, a possible solution is to perform a globally the liver to a whole CT scan volume. If the subset is too small optimised initial alignment. Given the topology of the liver and contains for example an individual vessel segment, the vascular tree, we hypothesise that it is possible to globally registration problem is poorly constrained and depends on a align a small LUS volume to the large CT volume if the vas- very accurate initialisation. However, our hypothesis is that culature captured in the LUS is geometrically unique. The given enough ultrasound data, a set of vessels and their rel- remaining challenge then becomes knowing in which areas ative pose and shape is likely to have a unique solution. The of the liver we can find such subsets. Therefore, we propose task of this paper is to determine what volume needs to be a pre-operative planning framework to tell the surgeon both scanned in order to obtain a unique vascular configuration as where these regions are and how much vascular data in their a function of location across the liver. This information can vicinity is required for a unique alignment. then be used to guide the surgeon to the appropriate liver vol- ume that needs to be acquired in order to achieve a reliable Contributions of this paper registration. To the best of our knowledge, there is currently no method to Background plan LUS acquisition in terms of registration outcome. Our contributions include: The Ultrasound (US) to CT registration problem has been tackled by several groups. Solutions for freehand US reg- – A comparison study on the assessment of which liver vascular features are best for registration. istration have been proposed either by matching US to a combination of the CT signal and a simulated US signal [4], – A framework that predicts how much data must be col- or by registering vessel probability maps from both US and lected at each region of the liver for a globally optimal CT [5]. These methods have been demonstrated using images registration. that capture large sections of liver and our preliminary tests – A proposed global registration approach for LUS to CT suggest such intensity based methods do not work well on the data, with results that are compared to the planning. restricted volume captured by the LUS probe. Other authors aligned 3D US instead, using reconstructed vessels as fea- tures [6,7], vessels and liver surface [8,9] or image intensity Vascular features assessment gradient information [10]. However, there are currently no 3D probes for laparoscopy on the market. In order to understand the nature of the vascular data that is to Few authors have investigated the specific registration of be used in our planning framework, we perform a preliminary LUS to CT. The first feasibility studies were proposed by Bao study that compares the registration outcome when using two et al. in an isolated phantom [11] and Kruecker et al. [12]on distinct vascular features: the vessel branching points, also a complete experimental setup. The first animal results in know as bifurcations, and the vessel centreline points. an ex-vivo porcine liver were reported by Martens et al. [13] For this purpose, we simulate the perfect acquisition of using a surface-based rigid registration with the LapAssistent subsets of N bifurcations and surrounding neighbouring system. During surgery, besides being moved by breathing centrelines, apply displacements to simulate expected defor- motion, the liver is compressed by the LUS probe upon mations or segmentation inaccuracies, and register them back contact and deformed by pneumoperitoneum (abdominal to the total model (see Fig. 1). Bifurcations are registered insufflation required for laparoscopy). This makes globally using point-based registration [17], mimicking the scenario rigid solutions such as the three aforementioned ones not where a surgeon picks common landmarks in both modalities sufficient. Even though deformable solutions are available and performs a rigid alignment. Centrelines are registered [7,8,10,14,15], they are hard to validate clinically and may using Iterative-Closest-Point (ICP) [18,19], simulating the require unfeasible computation times during surgery. Song case where correspondences are not known, and the best et al. [1] proposed locally rigid registration to be sufficient rigid solution using segmented centrelines is computed. In given a small enough liver region of interest, and validated the case where there are no displacements, the resulting trans- a vessel-based approach on in-vivo porcine data. However, formation is the identity. Otherwise, a new transformation is the method required manual selection and matching of vessel obtained and applied to the whole liver model. Using the bifurcations between LUS and CT, a task that is very diffi- new bifurcation positions and the original ones, we compute cult and time-consuming during surgery. Previously, in [16], a simulated Target Registration Error (TRE) over all bifur- we attempted vessel-based registration without establishing cations and assess the performance of both approaches. landmark correspondences, but the results still depended on We use this method to test the influence of the number of the accuracy of the initialisation. picked bifurcations and amplitude of applied displacements. 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1179 Fig. 1 Assessment of registration based on the two different vascular right are the resulting configuration, green the original one. Colormaps features. Point-based registration on bifurcations and ICP on centrelines show predicted displacements between these two, values ranging from are used to recover a transformation T . The red vessel models on the blue (low) to red (high) 50 combinations of N landmarks are randomly selected for N surrounding 3 bifurcations, and without specifying point cor- ranging from 3 to 10. For each combination, registrations are respondences. This indicates that within a liver region, fewer computed after applying a set of 1000 Gaussian distributed vascular branchings are required by the centreline ICP to displacements to the selected features, 200 for each devia- yield an accurate result. Since in a real scenario LUS reg- tion σ ranging from 1 to 5 mm. In the bifurcation approach, istration to CT needs to be accurate in a local region of displacements are applied in each point separately. For the interest, we also assess the accuracies in terms of a confi- centrelines, to better simulate deformation between branches, dence margin. For each registration, we compute the centre a random displacement vector is applied to all the points that of the picked points and find the maximum distance from it compose a branch. below which we capture bifurcations with TRE below 5 mm. The results of this experiment are displayed in Fig. 2. Each In the bottom row of Fig. 2, we present the mean results of curve describes the TRE results for a fixed number of bifur- this radial distance the same way as in the accuracy plots. In cations given different amplitudes of displacements. In the both methods, an increasing number of bifurcations increases top row, the Root Mean Square (RMS) of the bifurcation this distance. However, unless deviations surpass 3 mm, the TRE over all the vascular tree is presented in millimetres. confidence margins of ICP are superior to point-based reg- Each point in a curve averages a result of 200 registra- istration: ICP with centrelines around only 3 bifurcations tions over 50 bifurcation configurations. As expected, in both gives a wider mean working radius than point-based regis- methods the more bifurcations used, the lower the resulting tration on 5 bifurcations. Picking the previous example of 6 error. Considering 5 mm to be a clinically relevant level of landmarks and 1 mm of deviation, there is an improvement of a accuracy, we can see that using point-based registration 110–150 mm of the working radius. Therefore, besides being would require manual identification of 6 bifurcations with more accurate, centreline ICP also requires fewer branching an accuracy of 1 mm. This is observed in the green curve points to be reliable. with displacement σ = 1 mm. However, to obtain the same These results suggest that up to 6 landmarks would be result, the ICP approach would only require the centrelines required for a point-based registration. Clinically, identify- 123 1180 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 Fig. 2 Accuracy results on vessel bifurcations after running results of than 5 mm. The left column refers to bifurcation based registration; the the first experiment. The top row presents the RMS of the TRE on bifur- right column refers to centreline-based registration cations; the bottom row represents the radius where accuracy is better ing this number of landmarks during surgery would be very Uniqueness assessment challenging and time-consuming for a surgeon, making the vessel centreline approach preferable. We therefore conclude Our uniqueness assessment strategy relies on the use of any that vessel centrelines will provide a more clinically relevant Global ICP method, such as the Globally Optimal ICP (GO- registration result than bifurcations. ICP) proposed by Yang et al. [20]. This algorithm searches the solution space of rotations and translations with a Branch and Bound approach to find the rigid alignment that best minimises the L norm between two point clouds given a Methods minimum error threshold. We hypothesise that a subset of vessel centreline points is unique if GO-ICP aligns them cor- Given the demonstration above that centrelines should pro- rectly with the original complete tree using a sufficiently vide better registration results than bifurcations, our frame- small error threshold. Since the problem is rigid, the global work aims to assess in a planning stage which centreline point minimum of the distance function is 0, in which the subset sets are unique and possible to be aligned globally without is perfectly aligned in the original position. However, by set- initialisation. This assessment is performed over the liver sur- ting the threshold limit to a higher value, we allow GO-ICP to face, guiding a surgeon to acquire the necessary amount of find solutions with larger distance error. When the algorithm ultrasound data to acquire a unique set of vessels. 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1181 Fig. 3 Framework for assessment of uniqueness radius in a surface point of the pre-operative CT. Green represents points captured by a sphere around a point sampled on the pre-operative surface. Red represents GO-ICP resulting alignment reaches this value and outputs a configuration that is not close delineated in order to obtain binary masks of vessel sections; to the original one, the distance function has at least one local secondly, centreline points are estimated as the centroids of minimum that is not related to the global one, implying that each of these vessel sections. This segmentation strategy aims the subset is not geometrically unique. to simulate results that could be obtained with an automatic framework using only intensity information from 2D images. Currently, such approaches could be implemented for exam- Uniqueness planning ple using Deep Learning methods. An intra-operative probe contact position P of the dataset is computed as the mean Instead of using complete centreline branches as in “Vascular of the tracked surface digitised points as defined in [16]. features assessment” section, our framework evaluates the The same expanding radius strategy of “Uniqueness plan- uniqueness of centreline point subsets captured by simulated ning” section is then applied to the segmented centrelines probe positions along the pre-operative model surface. This using P as the sphere centre. In each registration, the TRE process is illustrated in Fig. 3. Firstly, the surface subset that is anterior and visible during surgery is extracted manually between manually picked bifurcations from both the LUS dataset and CT is measured. We define the success radius R and subsampled to generate possible probe contact positions. as the radius from which the TRE result obtained by GO- A sequence of spheres with increasing radius is generated ICP reaches a minimum and stabilises. If this minimum is around each of these positions, and the centreline points that not reached, we consider the registration to be unsuccess- are captured in its domain are tested for uniqueness with ful. GO-ICP. We then define the uniqueness radius R as the In the real LUS data several deformation components minimum value at which the captured centreline points are are expected. This means that the minimum L error found unique. By repeating this process in all the sampled points, a by GO-ICP may not be physically correct. To mitigate this R map is generated which tells the surgeon how broad the effect, we use some prior knowledge of the LUS acquisition acquisition must be in each liver region for a globally unique to constrain GO-ICP solution search space. For translations, registration. we constrain the dataset to the liver lobe where it was cap- tured, left or right. For rotations, we define a rough estimate LUStoCTregistration of the direction in which the sweep was taken in the CT, and ◦ ◦ allow only for rotations between [− 60 , 60 ]. This strategy To validate the uniqueness planning, we register continu- prevents the algorithm from considering physically impossi- ous sets of 2D tracked LUS images to the pre-operative CT ble acquisition configurations. using the same GO-ICP methodology of the previous sec- To compare the result with the planning, we obtain a tion. For each set, vessel centrelines are extracted from the reference standard transformation by applying point-based images in two steps: firstly, vascular structures are manually 123 1182 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 Fig. 4 Uniqueness map obtained by sampling 400 points over the visible surface. Sphere radius ranges from 5 to 100 mm in steps of 5 mm. Red and blue represent high and low values of uniqueness radius respectively registration in the landmarks used for TRE assessment. The face results to the whole visible surface points to obtain a aim here is to test how well our uniqueness map predicts complete uniqueness radius map. the amount of acquired data required for a reliable registra- Global registration experiments are performed in 5 sets of tion. Therefore, the radius R at which GO-ICP successfully LUS images, 3 acquired from the right lobe of the subject aligns the LUS data to the CT should be similar to the radius and 2 acquired from the left. In order to reduce fluctua- R predicted by the framework given the point-based refer- tion effects from the EM tracking system, we fit a cubic ence standard alignment. spline to the digitised surface and apply the result to the centreline points, smoothing them. We run the uniqueness assessment with the same parameters as the planning plus Experiments the prior solution space constraints described in “LUS to CT registration” section. Reference standard point-based align- Experiments were performed in retrospective data from a ments are obtained separately for each dataset. By applying single porcine subject previously acquired in [1]. Surface and the resulting transformation to the respective positions P, vascular models were segmented from standard abdominal reference probe positions P are obtained and used for com- triphase CT scans with resolution 512 × 512 × 2.5 mm. LUS parison with the generated R map. images of resolution 384 × 456 mm were acquired at a rate 2 3 of 10Hz using an Analogic SonixMDP and a Vermon LP7 linear probe which was electromagnetically (EM) tracked by Results an NDI Aurora tracking system at a rate of 40Hz. We test the planning framework by sampling 400 probe Uniqueness planning contact positions evenly spaced along the surface using the Farthest-point sampling algorithm [21]. We vary the sphere Figure 4 shows the obtained uniqueness radius map. Red radius in steps of 5 mm in the range [5–100] mm. We set the areas have a high uniqueness radius, implying that broader minimum threshold of GO-ICP to a mean distance error of data collections around them are required for a unique reg- 2.4 mm. Smaller values give finer results but require larger istration. In contrast, blue regions have a lower uniqueness computation times. We linearly interpolate the sampled sur- radius and require less data. The highest values are obtained at the liver edges, which is expected given that vasculature is less present in the liver periphery. The best and lowest val- www.visiblepatient.com. ues are observed in more central regions of the liver, ranging www.analogicultrasound.com. from 25 to 40 mm. The exception is the liver top surface www.vermon.com. which is more distant from the vasculature, yielding results www.ndigital.com. around 50 mm. 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1183 Fig. 5 Accuracy results of LUS to CT global registration experiments. Left represents results using a normal CT scan as target; Right represents results using CT scan under insufflation. Yellow markers represent the stable minimum that is measured as R LUStoCTregistration registration experiments using an insufflated CT vascular tree and therefore compensate for pneumoperitoneum. CT scans LUS to CT registration accuracy results are displayed in under these conditions are not acquired clinically, and hence the left plot of Fig. 5. For each dataset, the RMS of TRE are not used for planning. However, we present these results is presented as a function of the capture radius around the for illustrative purposes. Accuracy results of this experiment respective intra-operative probe contact position. For the are presented in the right-hand side of Fig 5. along with cases where TRE stabilises at a minimum and registration the respective radius measurements in the right-hand side of is successful, a yellow marker highlights the corresponding Table 1. In this case, all datasets are registered successfully success radius R . Table 1 summarises the obtained success with a R radius similar to the planning, with differences S S radii R along with the corresponding R predicted by the ranging from 0.5 to 9.1 mm. Furthermore, in agreement with S U planning framework. For each dataset, this R value is taken the successful cases of the normal experiments, the distances as the uniqueness radius measured at the planning surface D never surpass 16 mm. point closest to the reference probe position P . Error E Visually, we can observe this improvement in Fig. 6.In R S represents the RMS of TRE measured at radius R .Error E each row, the result of each of the three registration exper- S R represents the RMS of TRE obtained by the reference stan- iments is presented. We can observe that for Dataset 5, dard. Since this standard is the result of direct minimisation GO-ICP with the measured R of 30 mm aligns the recon- of the distance between the landmarks with a rigid solution, structed ultrasound sweep in the same region as the reference E provides the minimum TRE error that can be achieved. standard. The same happens if the insufflated scan is used, It is possible to see that 3 out of 5 cases were successfully with the R of 40 mm. In the case of Dataset 4, we observe registered, and that the respective R values are similar to that registration to the normal scan is not successful with the prediction R , with differences of 1.5, 7.2 and 5 mm. In any radius, but with an insufflated scan results improve and the two failure cases, registration is never successful, and no a radius R of 40 mm is observed. The same improvement comparison with planning can be made. To better understand from normal to insufflated scan was obtained with Dataset 3. these cases, we measure D , the distance between P and R R corresponding R planned location. In the successful cases, this distance does not surpass 15 mm. However, in the failure Discussion cases, this result surpasses 23 mm, indicating a large differ- ence between intra and pre-operative surface and therefore The results of our uniqueness mapping are as expected, since implying the presence of significant liver tissue deformation. surface regions that are far from the vasculature are predicted In order to assess how much deformation is influencing to require broader data acquisitions. In practical terms, the this registration problem, we repeat the same planning and mapping of Fig. 4 would guide a surgeon to acquire data 123 1184 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 Table 1 Success radius Dataset Normal CT scan CT scan under insufflation measurement results of LUS R R E E D R R E E D U S S R R U S S R R datasets 1 (Left lobe) 31.5 30.0 16.3 5.94.8 39.1 30.0 20.98.613.2 2 (Left lobe) 32.2 25.0 10.4 1.814.8 40.0 35.0 21.82.82.4 3 (Right lobe) 45.0 – – 4.725.0 43.8 40.0 9.66.815.8 4 (Right lobe) 40.0 – – 21.323.1 40.5 40.0 15.310.314.4 5 (Right lobe) 35.0 30.0 12.8 9.413.2 45.0 40.0 5.23.29.1 E represents the RMS of TRE obtained with registration at success radius R ; E represents the same error S S R for the reference standard transformation; R represents the uniqueness radius predicted by the framework using the reference standard; D represents distance between the reference probe position P and surface R R location of R measurement. All values are in millimetres Fig. 6 Visual results of registrations of LUS datasets 4 and 5 from the GO-ICP registration failure using maximum radius; Right shows GO- right lobe. Left shows the reference standard results. Middle top shows ICP result with insufflated scan as target at the measured success radius GO-ICP result with measured success radius; Middle bottom shows preferentially in the blue areas, where smaller volumes of cies ranging from 5.2 to 21.8 mm. Higher errors above 20 mm data are required for a globally optimal alignment. and much larger than the reference standard are observed in In the LUS to CT registration experiments, 3 out of the left lobe datasets. This may be explained by errors in 5 datasets were registered successfully with accuracy val- the manual localisation of bifurcations. Furthermore, retro- ues E of 16.3, 10.4 and 12.8 mm (Table 1 column 3). spective analysis showed that the CT insufflated scan was In the two failure cases, large differences between intra- unreliably segmented in this region. Regardless, the datasets operative and pre-operative surface positions indicate defor- are aligned in the correct region, but with a larger translation mation. From the three expected deformation factors listed error. These results indicate that this registration problem in “Background” section, surface compression and pneu- may be more affected by insufflation than surface compres- moperitoneum are the ones that may better explain this sion. phenomenon. Further experiments with an insufflated target In the normal scan experiments, for the 3 successful cases, led to successful registrations with all datasets with accura- the planned radius R is similar to the measured success 123 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 1185 radius R with a maximum difference of 7.2 mm. In the Acknowledgements JR, MRR were supported by the EPSRC CDT in Medical Imaging [EP/L016478/1]. MJC, DH, DB, ST were supported insufflated case, a larger difference of 9.1 mm is observed for by the Wellcome/EPSRC [203145Z/16/Z]. BD, DH was supported by dataset 1, a fact which is possibly related to the lower accu- the NIHR Biomedical Research Centre at University College London racy of the registration. However, it can be observed that Hospitals NHS Foundations Trust and University College London. The the radii R are always lower than the planned R . Since imaging data used for this work was obtained with funding from the S U Health Innovation Challenge Fund (HICF-T4-317), a parallel funding we corrected for insufflation in the second experiment, this partnership between the Wellcome Trust and the Department of Health. trend could be related to surface compression—if the surface The views expressed in this publication are those of the author(s) and not is pushed closer to the vessels, unique vasculature would be necessarily those of the Wellcome Trust or the Department of Health. found with a lower capture radius. Regardless, the difference between predictions and measurement is not critical as a sur- Compliance with ethical standards geon would not be able to collect data with a 5 mm radius precision. In addition to showing a reasonable agreement with the Conflict of interest The authors declare that they have no conflict of planning predictions and the obtained results, this approach interest. points to the feasibility of GO-ICP method with prior knowl- Ethical approval All applicable international, national, and/or institu- edge for vessel-based registration of LUS to CT. The obtained tional guidelines for the care and use of animals were followed. All TRE values obtained in the normal experiment are promis- procedures performed in studies involving animals were in accordance ing given the fact that we aimed to initially align rigidly a set with the ethical standards of the institution or practice at which the studies were conducted. For this type of study, formal consent is not of vessels in its correct region. Obtaining clinically valuable required. This article does not contain patient data. accuracies below 10 mm could be achieved by compensat- ing for deformation locally afterwards. Using non-optimised Open Access This article is distributed under the terms of the Creative code and hardware, the uniqueness map compilation and Commons Attribution 4.0 International License (http://creativecomm ons.org/licenses/by/4.0/), which permits unrestricted use, distribution, clinical registration with GO-ICP took 20 h and 10 min and reproduction in any medium, provided you give appropriate credit respectively. Since the former is a pre-operative off-line step, to the original author(s) and the source, provide a link to the Creative this high value would not be critical to clinical application. Commons license, and indicate if changes were made. Conclusions References We have developed a pre-operative planning framework to 1. Song Y, Totz J, Thompson S, Johnsen S, Barratt D, Schneider guide the surgeon in acquiring LUS data in a way that a C, Gurusamy K, Davidson B, Ourselin S, Hawkes DJ, Clarkson reliable registration to CT can be obtained without an initial MJ (2015) Locally rigid, vessel-based registration for laparoscopic manual alignment. We consider this framework to be relevant liver surgery. In: IJCARS, pp 1–11 2. Ciria R, Cherqui D, Geller DA, Briceno J, Wakabayashi G (2016) since LUS is particularly difficult to handle, and providing Comparative short-term benefits of laparoscopic liver resection: guidance beforehand could increase the use of this regis- 9000 cases and climbing. Ann Surg 263(4):761–777 tration method and enhance the safety of laparoscopic liver 3. Wakabayashi G, Cherqui D, Geller DA, Buell JF, Kaneko H, Han resection. HS, Asbun H, O’Rourke N, Tanabe M, Koffron AJ (2015) Rec- ommendations for laparoscopic liver resection: a report from the The major limitation of this work is the restriction of the second international consensus conference held in morioka. Ann study to one subject. To mitigate this risk, we show variabil- Surg 261(4):619–629 ity by applying our method on data from several different 4. Wein W, Brunke S, Khamene A, Callstrom MR, Navab N (2008) regions of the liver. Another limitation maybe the fact that Automatic CT-ultrasound registration for diagnostic imaging and image-guided intervention. Med Image Anal 12(5):577–585 we do not compensate for any deformation in the framework, 5. Penney GP, Blackall JM, Hamady MS, Sabharwal T, Adam A, leading to registration failures. However, for the purpose of Hawkes DJ (2004) Registration of freehand 3D ultrasound and a globally optimal initial alignment, it is difficult to intro- magnetic resonance liver images. Med Image Anal 8(1):81–91 duce deformation parameters. Since successful registrations 6. Aylward SR, Jomier J, Guyon JP, Weeks S (2002) Intra-operative 3D ultrasound augmentation. In: ISBI, pp 421–424 were achieved with the insufflated scan, rigid may be suffi- 7. Lange T, Eulenstein S, Hünerbein M, Schlag PM (2003) Vessel- cient if insufflation is modelled in the pre-operative CT, as in based non-rigid registration of MR/CT and 3D ultrasound for [22]. navigation in liver surgery. Comput. Aided Surg. 8(5):228–240 Our future work directions are twofold. From a techni- 8. Lee Duhgoon, Nam Woo Hyun, Lee Jae Young, Ra Jong Beom (2011) Non-rigid registration between 3d ultrasound and ct images cal point of view, we intend to simulate insufflation in our of the liver based on intensity and gradient information. Phys Med framework and in a later stage probe compression. In terms Biol 56(1):117 of validation, we aim to apply this framework in human data 9. Nam WH, Kang DG, Lee D, Lee JY, Ra JB (2012) Automatic reg- in order to obtain clinically translatable results. istration between 3d intra-operative ultrasound and pre-operative 123 1186 International Journal of Computer Assisted Radiology and Surgery (2018) 13:1177–1186 ct images of the liver based on robust edge matching. Phys Med 17. Arun KS, Huang TS, Blostein SD (1987) Least-squares fitting of Biol 57(1):69 two 3-D point sets. IEEE Trans Pattern Anal Mach Intell 5:698–700 10. Lange T, Papenberg N, Heldmann S, Modersitzki J, Fischer B, 18. Besl PJ, McKay ND (1992) Method for registration of 3D shapes. Lamecker H, Schlag PM (2009) 3D ultrasound-CT registration of In: Proceedings of SPIE 1611, Sensor fusion IV: control paradigms the liver using combined landmark-intensity information. IJCARS and data structures, pp 586–606 4(1):79–88 19. Chen Y, Medioni G (1991) Object modeling by registration of mul- 11. Bao P, Warmath J, Galloway R, Herline A (2005) Ultrasound-to- tiple range images. In: IEEE international conference on robotics computer-tomography registration for image-guided laparoscopic and automation, pp 2724–2729 liver surgery. Surg Endosc 19(3):424–429 20. Yang J, Li H, Campbell D, Jia Y (2015) GO-ICP: A globally optimal 12. Kruecker J, Viswanathan A, Borgert J, Glossop N, Yang Y, Wood solution to 3D ICP point-set registration. IEEE Trans Pattern Anal BJ (2005) An electro-magnetically tracked laparoscopic ultra- Mach Intell 38(1):12 sound for multi-modality minimally invasive surgery. Int Congr 21. Peyré G (2011) The numerical tours of signal processing. Adv Ser CARS 1281:746–751 Comput Signal Image Process Comput Sci Eng 13(4):94–97 13. Martens V, Besirevic A, Shahin O, Kleemann M (2010) Lapas- 22. Johnsen SF, Thompson S, Clarkson MJ, Modat M, Song Y, Totz J, sistent: computer assisted laparoscopic liver surgery. In: IFMBE, Gurusamy K, Davidson B, Taylor ZA, Hawkes DJ (2015) Database- vol 22 based estimation of liver deformation under pneumoperitoneum for 14. Dagon B, Baur C, Bettschart V (2008) Real-time update of 3D surgical image-guidance and simulation. In: MICCAI. Springer, pp deformable models for computer aided liver surgery. In: ICPR, pp 450–458 1–4 15. King AP, Blackall JM, Penney GP, Hawkes DJ (2001) Tracking liver motion using 3D ultrasound and a surface based statistical shape model. In: MMBIA 2001, pp 145–152 16. Ramalhinho J, Robu M, Thompson S, Edwards P, Schneider C, Gurusamy K, Hawkes DJ, Davidson B, Barratt D, Clarkson MJ (2017) Breathing motion compensated registration of laparoscopic liver ultrasound to CT. In: Proceedings of SPIE medical imaging

Journal

International Journal of Computer Assisted Radiology and SurgerySpringer Journals

Published: Jun 2, 2018

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off