Comparison study of intraoperative surface acquisition methods on registration accuracy for soft-tissue surgical navigation

Abstract.

Purpose

To study the difference between rigid registration and nonrigid registration using two forms of digitization (contact and noncontact) in human in vivo liver surgery.

Approach

A Conoprobe device attachment and sterilization process was developed to enable prospective noncontact intraoperative acquisition of organ surface data in the operating room (OR). The noncontact Conoprobe digitization method was compared against stylus-based acquisition in the context of image-to-physical registration for image-guided surgical navigation. Data from $n=10$ patients undergoing liver resection were analyzed under an Institutional Review Board-approved study at Memorial Sloan Kettering Cancer Center. Organ surface coverage of each surface acquisition method was compared. Registration accuracies resulting from the acquisition techniques were compared for (1) rigid registration method (RRM), (2) model-based nonrigid registration method (NRM) using surface data only, and (3) NRM with one subsurface feature (vena cava) from tracked intraoperative ultrasound (NRM-VC). Novel vessel centerline and tumor targets were segmented and compared to their registered preoperative counterparts for accuracy validation.

Results

Surface data coverage collected by stylus and Conoprobe were $24.6\%\pm 6.4\%$ and $19.6\%\pm 5.0\%$, respectively. The average difference between stylus data and Conoprobe data using NRM was $-1.05\mathrm{mm}$ and using NRM-VC was $-1.42\mathrm{mm}$, indicating the registrations to Conoprobe data performed worse than to stylus data with both NRM approaches. However, using the stylus and Conoprobe acquisition methods led to significant improvement of NRM-VC over RRM by average differences of 4.48 and 3.66 mm, respectively.

Conclusion

The first use of a sterile-field amenable Conoprobe surface acquisition strategy in the OR is reported for open liver surgery. Under clinical conditions, the nonrigid registration significantly outperformed standard-of-care rigid registration, and acquisition by contact-based stylus and noncontact-based Conoprobe produced similar registration results. The accuracy benefits of noncontact surface acquisition with a Conoprobe are likely obscured by inferior data coverage and intrinsic noise within acquisition systems.

1. Introduction

Each year, primary liver cancer arises in $\sim \mathrm{41,000}$ people, and there are more than 30,000 related deaths in the United States.¹ Worldwide, primary liver cancer is responsible for $\sim \mathrm{900,000}$ new cases per year and 830,000 related deaths.¹ Further compounding to overall incidence, the liver is a frequent site for metastatic disease from other primary cancer sites. For example, up to 50% of colorectal cancer patients will have cancers that metastasize to liver. Surgical treatments are a viable curative option for isolated metastases, meaning that upon detection of metastasis, the lesions are only detected in the liver.²

Liver resection is considered the gold standard approach to treating selected patients with primary and metastatic liver disease due to its excellent long-term survival outcomes. Given the regenerative properties of the liver, liberal margins are often possible for liver resection performed with curative intent. However, precision is still needed to preserve vascular and bile duct integrity as well as maintain appropriate healthy liver remnant. Image-guided liver surgery (IGLS) systems in the development are a potentially important tool for liver surgery that could provide detailed preoperative information on the patient-specific liver anatomy, extent of disease, and estimated remnant liver volume to the surgeon. These IGLS systems have the ability to co-locate this information on the physical patient while surgery proceeds. Yet, intraoperative deformation of the liver is unavoidable due to procedural aspects. For example, in open approaches, perihepatic packing, tissue retraction, and organ mobilization impart intraprocedural deformations compared to preoperative imaging that can compromise guidance fidelity. Systems to compensate for such deformations are therefore critical. The scale of these deformations can be considerable. In open surgery, a previous study was conducted that involved rigidly registering each patient’s preoperatively liver surface to its digitized intraoperative counterpart.³ When compared, it was found that the signed closest point distance between registered surfaces varied as much as $-13.4$ to 16.2 mm on average. These deformations incur a fundamental mismatch between the physical organ in its intraprocedural conformation and its preoperatively imaged counterpart. Several approaches for correcting deformation have been developed for image-to-physical registration using partial liver surfaces. In this work, the linearized iterative boundary reconstruction (LIBR) nonrigid registration method (NRM) proposed by Heiselman et al.⁴ is utilized.

The objective of this study is to investigate the quantitative differences between rigid registration and NRMs with respect to two forms of digitization, which are contact and non-contact digitization methods. To date, all FDA-cleared approaches for IGLS have involved intraoperative acquisition of localization data using an optically tracked stylus that makes physical contact with the organ surface during measurement. While surface data measurement with a tracked pointer is appropriate for rigid structures that do not deform, in soft-tissue applications, registration accuracy can be compromised if the pointer loses contact with the surface or if the tissue deforms under the stylus due to contact forces. A previous study demonstrated that noncontact digitization methods outperform the contact method in phantom experiments, ex vivo liver experiments, and ex vivo kidney experiments. In those examples, target registration error (TRE) with the noncontact method was reduced by more than 2 mm over its contact-based counterpart.⁵ However, these highly controlled environments are not representative of the operating room (OR). In this work, we explore the use of these methods under true surgical conditions.

2. Method

2.1. Surface Data Acquisition Methods

2.1.1. Conoprobe attachment design

Our noncontact digitization method utilizes an optically tracked Conoprobe Mark 10 device (Optimet Metrology Ltd., Jerusalem, Israel), which uses polarized light interference inside of a birefringent crystal to acquire surface range data. During a measurement, the target point is illuminated with a columnated light source such as a laser and then projected along the outward optical path via a beam splitter. The uncolumnated reflected light returning to the Conoprobe passes through a polarizing filter and then a birefringent crystal which generates an interference pattern. The pattern is projected through a second polarizing filter and then recorded using a charge-coupled device sensor. Distance measurements can be computed from the resulting Fresnel zone pattern.⁶ To translate this surface measurement method into the OR, the first challenge was the realization of a handheld Conoprobe device compatible with the sterile environment. A novel device attachment and workflow design shown in Fig. 1 was developed to enable OR use. The attachment was designed using a CAD software (Autodesk Fusion 360, San Francisco, California, United States). Of note, the offset for the fiducial markers needed to have sufficient clearance such that hand-held operation was enabled without disturbing tracking. Aluminum was selected as a lightweight material for construction of the sterilizable attachment. An aperture window was built into the attachment design to prevent nonsterile debris material from escaping into the sterile environment. With the customized attachment design, a calibration and validation process was required to relate the Conoprobe measurement with the tool tracking coordinate frame.

Fig. 1.

(a) CAD model of the design, (b) Conoprobe with attachment design, and (c) Conoprobe use in OR with sterile bag.

Figure 2 illustrates the construction of tracking frame coordinates. The Conoprobe is equipped with optical tracking markers within which a tool coordinate frame $C$ is assigned. The Conoprobe returns a measurement distance $D$ performed with respect to the origin of the device lens frame $L$. This distance measurement is associated with an unknown laser beam measurement direction $d$ and lens offset vector $\mathbf{o}$ associated with the lens frame $L$ in $C$. A measured point $p$, therefore, can be expressed in the local coordinate frame $C$ as $p=o+Dd$. To determine the calibration parameters o and d, a point is measured by an optically tracked stylus to determine its ground truth and then repeatedly measured seven times by the Conoprobe at different distances and poses which is $q_i$. For each Conoprobe measurement, the distance $D$ and the rigid body transformation $T_{\mathrm{RB}}$ from the coordinate frame $C$ to the ground-truth tracking frame $W$ are recorded. The relationship is shown in Fig. 2. The vector $\mathbf{o}$ and unit vector $\mathbf{d}$ can be calculated by least-squares fitting.⁷ The quality of the calibration is determined by the calibration residual error, which is the root-mean-square error between $q_i$ and $p_i$. The accuracy of the tracked Conoprobe for point measurements was then validated using a stair block phantom. The stair block phantom was designed specifically with a fiducial point in the center of each red disk of the nine black platforms with each platform having varying heights and each platform consists of centered disks with a 3 mm hemispherical divot.⁷ As a comparator in digitization, the ground-truth position of the fiducials was measured by an optically tracked stylus. For each dataset, the nine points were then measured again with the tracked Conoprobe. A total of $n=9$ datasets were acquired where the pose of the optical tracking system was changed for each dataset. While altering the optical tracker pose between the acquisition of the datasets, the step phantom is static relative to the tracking cameras; hence, all measurements are in the same frame, i.e., are registered to each other. This process allows direct comparison between the points measured with the Conoprobe and the ground-truth points provided by the tracked stylus. The average of the Euclidean distance error between these two sets of points was calculated to obtain the calibration TRE. In addition, the attachment described in Fig. 1 will be used repeatedly in the OR and it is desirable to only calibrate the Conoprobe once. The potential impact of reattachment reproducibility on accuracy was therefore also characterized. After calibration, a series of $n=10$ detach and reattach cycles of the Conoprobe tracking piece were performed. For each reattachment, $n=9$ points were acquired using the optically tracked stylus on the stair block as their ground-truth position, and the same points were measured by the Conoprobe with the reattached tracking piece to redetermine calibration TRE.

Fig. 2.

Conoprobe calibration. The calibration parameters are the offset vector $\mathbf{o}$ to the lens frame $L$ and the laser direction $\mathbf{d}$ in the local coordinate frame $C$ of the sensor.

2.1.2. Previous closely related work

Prior to moving on to the structure of the clinical study performed in this work, it is important to recognize previous findings that are closely related to this work. As noted in the introduction, the impetus of this work centers on the potential benefits of noncontact surface digitization with the Conoprobe over the contact-based tracked stylus swabbed digitization. In the previous work,⁵ a series of alignment comparisons were performed with soft tissue organs (ex vivo porcine liver and human cadaveric kidney). The surface registration errors reported using the Conoprobe were $1.73\pm 0.77$ and $1.50\pm 0.50\mathrm{mm}$ for the liver and kidney specimens, respectively, whereas the swabbed counterparts were $4.44\pm 1.19$ and $3.51\pm 0.82\mathrm{mm}$. In repeated testing, the benefit of noncontact digitization was statistically significant. However, as also noted above, the testing conditions in these experiments, while still using soft-tissue organs, were quite idealized in control and exposure. Related to this last point regarding exposure, i.e., available surface area for registration, in other related work⁸ that analyzed surface coverage and the benefit of nonrigid alignment approaches, there was performance improvement using nonrigid alignment approaches over rigid approaches when surface coverage exceeded $\sim 22\%$ of the surface area. As such, the results from the intraoperative use evaluated below need to be understood in the context of the anticipated performance originating from the more controlled setting as a benchmark for expectations.

2.2. Comparison of Digitization Methods

2.2.1. Clinical data collection and processing

This preliminary study involved $n=10$ patients undergoing liver surgery and enrolled at Memorial Sloan Kettering Cancer Center under an Institutional Review Board (IRB) approved protocol with informed consent of all participants. For each patient, four types of data were collected including (i) preoperative computed tomography/magnetic resonance imaging, (ii) intraoperative sparse liver surface data collected by optically tracked stylus, (iii) intraoperative sparse liver surface data collected by the Conoprobe equipped with the sterilizable OR tracking attachment, and (iv) tracked iUS images for the purposes of registration and validation. For the first type of data, preoperative three-dimensional (3D) liver models were generated using custom surgical planning software and a series of anatomical features including the left and right inferior ridges and falciform ligament were identified. Intraoperative organ surface data along corresponding anatomical features were also acquired for the purposes of driving the salient feature rigid registration.⁹ For Conoprobe data, the signal-to-noise ratio (SNR) was provided by the device for assessing data quality. In our experience, data quality could diminish when the device was too close or too far from the object of interest. In our clinical experience to date, the data collected by the Conoprobe has been found to suffer without careful consideration of the SNR. As a result, two filters were implemented to denoise the data. The first is an SNR filter. All the points whose SNR during measurement were below 15% were eliminated. The remaining points were then treated with a second K-closest points outlier filter.¹⁰ The filter computes the average and standard deviation of the distances between each point and its nearest $k$ fraction of neighbors. A distance threshold was calculated so that any point was rejected if its closest point distance exceeded two standard deviations away from the mean. With the original intraoperative surface data by stylus and the processed intraoperative surface data by Conoprobe, accuracy results from both rigid registration and nonrigid registration were compared for each measurement technique. Moreover, previous studies have shown that the surface coverage of intraoperative data could impact the registration results, therefore, for each intraoperative surface data measurement technique, the extent of organ surface coverage as the percentage of boundary nodes on the preoperative liver model contained within an alpha shape constructed around the sparse data was determined.⁸

2.2.2. Rigid registration method

The rigid registration method (RRM) involved two steps. The first step was a point-based registration of the centroids of anatomical features identified in both the preoperative data and the intraoperative digitized point clouds to provide an initial pose estimation. The second step was a surface-based registration that utilized a modified iterative closest point (weighted ICP) algorithm between the preoperative liver surface and the intraoperative point cloud⁹ to further refine the pose estimate. Weighted ICP was used due to its incorporation of salient anatomical features into a weighting scheme that improves the robustness of ICP pose estimation. With respect to the weighted ICP weighting factor and how features are treated, this algorithm biases the closest point operator toward preferential alignment of salient features by reducing the Euclidean distance between feature point pairs among the preoperative and intraoperative salient feature designations. Although this initial bias allows for robust alignment, as the ICP registration proceeds, these weights are scheduled to reduce and allow for nonweighted data to more greatly influence alignment.⁹

2.2.3. Nonrigid registration method

The NRM begins with the rigid registration above and then uses sparse intraoperative data measurements within the context of an optimization approach to reconstruct a combination of precomputed solutions from a suite of boundary conditions representing potential intraoperative liver deformations. This approach to nonrigid registration is realized via the LIBR algorithm⁴ employed herein. With respect to the biomechanics associated with LIBR, the liver is simulated as an isotropic linearly elastic material and model deformations are solved using the finite-element method. More specifically, at static equilibrium, linear elasticity is governed by the Navier–Cauchy equations in three dimensions:

$$\frac{E}{2(1+v)}{\nabla }^{2}u+\frac{E}{2(1+v)(1-v)}\nabla (\nabla ·u)+F=0$$ (1)

where $E$ is the Young’s modulus, $v$ is the Poisson ratio, $u$ is displacement, and $F$ is applied force. The values $E=2100\mathrm{Pa}$ and $v=0.45$ were used in this work.⁸ However, it is important to recognize that the LIBR method in this realization is a displacement-driven boundary reconstruction within the context of a homogeneous material representation of the liver. As such, the performance of the registration is only sensitive to Poisson’s ratio material property. In the preoperative computation phase, a series of control points are evenly distributed across the liver surface. These control points are then perturbed in three orthogonal directions and the resulting linear elastic responses based on Eq. (1) are used to establish a deformation basis for the organ geometry. A local neighborhood processing step is conducted to create realistic boundary conditions as part of the conditioning of these basis solutions. From these conditioned displacement responses, stress and strain responses are also calculated for use within a strain energy regularization component in a least-squared error optimization framework that is executed during surgery. In the intraoperative computation phase, a full distribution of registration parameters consisting of rigid 6-DOF components and a distribution of coefficients associated with the deformations derived from the control node perturbations is optimized to match model-driven shape change between the preoperative organ and the physically acquired organ data. This optimization is performed in the context of an objective function [Eq. (2)] designed to minimize the difference between predicted and acquired shape, where $f_i$ denotes the error between the deformed organ model and an intraoperative data point within an intraoperatively collected point cloud for feature $F$ of size $N_F$, $w_F$ is the weight of the feature, $f_E$ is the average strain energy of the deformation state, and $w_E$ is a regularizing strain energy weight that controls the deformability of the registration:

$$\mathrm{\Omega }(\beta )=\sum _F\frac{w_F}{N_F}\sum _{i=1}^{N_F}{f}_i^2+w_E{f}_E^2.$$ (2)

This objective function distinguishes the error terms for distinct types of features that comprise the intraoperative data. From organ surface data, the features include the falciform ligament, the left and right inferior ridges, and the general anterior liver surface. In this study, the potential of the vena cava centerline as an additional subsurface feature that could influence the registration, i.e., an additional salient feature to the inferior ridges and falciform was also investigated. For the nonrigid registration method incorporating the vena cava from tracked iUS (NRM-VC), the features in Eq. (2) also included the vena cava centerline associated with hepatic vein confluence. Both forms of surface data (contact-based stylus and noncontact Conoprobe) were used within the context of rigid registration and nonrigid registration and were systematically compared among intraoperatively acquired liver targets from tracked iUS.

2.3. Validation with Tracked iUS

In our liver navigation system prototype, sparse subsurface data were collected from tracked iUS imaging for use within registration and to establish measures of target error. The tracked iUS setup consisted of a T-probe transducer attached to a sterilized optically tracked reference target with the system being calibrated using the N-wire phantom method.¹¹ For iUS validation, the contour segmentation of a subsurface target (i.e., tumors and vessels) was performed by students who are familiar with the anatomical structures on the iUS images and then transformed by either the rigid or NRM.

The quantitative error metric used was the mean closest distance between the target contours segmented from iUS images and the 3D anatomical model of the corresponding structure.¹² As an improvement for the method proposed in Ref. 12, instead of comparing a segmented contour to the liver model, we propose a centerline-to-centerline approach as shown in Fig. 3. For each target, multiple contours are segmented on the tracked iUS images in which the specific target is visible. For vessel targets, the centerline of the target was extracted from the 3D liver model using the vascular modeling toolkit.¹³ Then $\sim 8$ contours were segmented on the iUS images, and a polynomial curve fitting model was applied through the center point of each contour to reconstruct the centerline of the target vessel. Therefore, correspondence could be assured during the contour selection process. The extracted centerline from the model and the reconstructed centerline from the iUS images were compared by computing the mean closest distance between centerline features. For tumor targets, more than 20 contours from iUS images were segmented in various image orientations to cover the entire tumor volume. The centroid of the intraoperative tumor contours were compared to the centroid of the preoperative tumor model.

Fig. 3.

Overview of the centerline-to-centerline approach. In case the target is vessel, several contours of a target are segmented from different iUS images to reconstruct a centerline of the vessel. The reconstructed centerline is then compared to the extracted centerline from the 3D model that is either preoperative or deformed. In case the target is a tumor, contours from all directions will be segmented to calculate the center point and then compared with the center point of the tumor model which is either preoperative or deformed.

An inherent problem with iUS validation is the shape distortion of visualized tissue structures introduced by the probe pressure on the organ surface during measurement and the incomplete/nonuniform sampling of structural shapes. Additionally, tissue compression during ultrasound imaging itself induces error in the location and geometry of subsurface features that may be used for registration or as targets. In this work, a modified version of the method introduced by Pheiffer and Miga¹⁴ to correct for the influence of probe deformation was implemented. The method was modified to only require a sparse intraoperative measurement of the liver surface. Intraoperative measurements of the noncompressed tissue surface using the digitization method (stylus or Conoprobe) were acquired, and then this surface was co-located with the position and orientation of the tracked ultrasound probe to estimate the depth to which the tissue was compressed. The intraoperative surface data were resampled using method proposed by Collins et al.¹⁵ to enrich the surface data, which could ensure that the surface data could be found in every ultrasound image. The surface data resampling method is to preprocess sparse data to normalize quality across acquisitions. After collection, a discretized surface is fit to the data using an interpolant to fit locally and surface Laplacian to smooth. The surface is then trimmed to the bounds of the initial data using a dilate and fill technique. The intersection points between the resampled surface data and the tracked iUS transducer position were computed via a ray-casting approach to determine the location where the iUS probe would be positioned without compression. The compression depth was then used to assign Dirichlet boundary conditions to a finite element mesh generated from the iUS image. These fixed displacement boundary conditions were employed at the iUS image-to-tissue interface, whereas the boundary conditions in the far-field were considered to be zero displacement. Lateral boundary conditions were considered stress-free, and conditions of plane strain mechanics were assumed for the mechanical model.¹⁶ After the assignment of boundary conditions, the model was solved for 2D displacements over the entire mesh to estimate the decompressed state of the tissue. Figure 4 shows an overview of each step in the compression compensation method.

Fig. 4.

Overview of the compression compensation method. For each iUS image, the segmented contour is compensated based on the displacement between the probe position and the intraoperative surface data.

3. Results

For the Conoprobe attachment experiment, the calibration residual error was $1.2\pm 0.4\mathrm{mm}$ and the calibration TRE was $1.7\pm 0.6\mathrm{mm}$. The TRE for the reattachment experiment was $1.8\pm 0.6\mathrm{mm}$. Additionally, for each patient, the surface extent percentage of intraoperative points collected by both stylus and Conoprobe are computed and compared in Table 1. On average, the organ surface coverage of stylus data was $24.6\%\pm 6.4\%$ and exceeded that of Conoprobe data ($p=0.025$, paired $t$-test), which was $19.6\%\pm 5.0\%$.

Table 1.

Summary of the coverage of the surface data collected by both stylus and Conoprobe. The surface coverage is the percent of the surface area that a registered point cloud covers as compared to the entire mesh surface.

Patient	Stylus (%)	Conoprobe (%)
1	31.7	24.6
2	23.2	14.5
3	15.7	15.6
4	32.9	28.6
5	31.7	16.4
6	22.3	16.0
7	20.9	13.1
8	13.5	20.5
9	26.3	22.0
10	28.1	25.3
Mean	24.6 ± 6.4	19.6 ± 5.0

Qualitative visualizations are presented to represent the impact of the NRM with vena cava on the alignment of the tracked iUS with the corresponding anatomical features generated from the preoperative tomograms. Figure 5 shows a rendering of the alignment of a tracked iUS image of the left portal vein confluence for patients 2, 6, and 3 under conditions of rigid and nonrigid registration. It should be noted that the visualization of the result of the nonrigid registration involves both the deformed anatomical models and an updated rigid body transform. The result shown in Fig. 5 represents an example of our best target alignment result based on stylus data (patient 2), the best target alignment result based on Conoprobe data (patient 6), and the worst overall target alignment result which happens to be based on the stylus data (patient 3).

Fig. 5.

Visualization of the tracked iUS image plane rendered within the 3D models generated from the preoperative tomograms for three patients based on tracked stylus and Conoprobe data. (a) The raw iUS image capture used in the analysis and (b) the highlighted anatomical feature (left portal vein confluence) used as validation target. (c), (d) The rigid registration transformation computed during the surgical procedure was used to generate the overlays. (c) A superior-view of the rigid alignment. (d) A zoomed superior-view of the tracked iUS and portal vein structure. (e), (f) The deformed anatomical models and updated transform from NRM with vena cava feature were used to generate the renderings. (e) The analogous superior-view of the deformation corrected alignment. (f) A zoomed superior-view of the tracked iUS and portal vein structure after deformation correction.

A summary of the centerline-to-centerline closest-point distance measurements for 30 anatomical subsurface targets across the 10 patients where registration was performed using rigid registration and nonrigid registration with/without vena cava feature, using intraoperative surface data acquired by “tracked stylus” and “tracked Conoprobe” are shown in Table 2, respectively. All the centerlines for US validation are generated by segmented contours applied with the compression compensation method. For both tracked stylus data and tracked Conoprobe data, nonrigid registration outperformed rigid registration in all 10 patients. For both stylus and Conoprobe data, adding the subsurface vena cava feature in the nonrigid registration exhibited a modest improvement in registration accuracy ($\sim 0.6$ and 0.1 mm, respectively).

Table 2.

Summary of the centerline-to-centerline closest-point distance errors based on tracked stylus and Conoprobe data between homologous features delineated in the preoperative tomograms and the tracked iUS image sets under conditions of rigid registration, nonrigid registration, and nonrigid registration with vena cava feature. The compensation method is applied in each method for more accurate targeting. S_means stylus data. C_means Conoprobe data.

Patient	Gender	Age	S_RRM error (mm)	C_RRM error (mm)	S_NRM error (mm)	C_NRM error (mm)	S_NRM-VC error (mm)	C_NRM-VC error (mm)
1	M	43	16.6 ± 1.7	17.6 ± 2.6	12.6 ± 2.0	13.4 ± 2.2	11.1 ± 1.2	12.8 ± 2.0
2	F	25	13.8 ± 1.3	16.3 ± 1.1	9.9 ± 1.0	12.9 ± 1.6	9.8 ± 1.4	12.4 ± 0.9
3	M	49	19.2 ± 2.9	17.4 ± 2.2	14.8 ± 2.2	15.1 ± 2.0	12.7 ± 1.8	12.2 ± 1.3
4	M	38	17.7 ± 3.0	18.4 ± 3.5	15.2 ± 1.9	14.5 ± 2.9	14.8 ± 2.7	14.4 ± 2.0
5	M	76	16.0 ± 3.2	16.9 ± 3.0	11.2 ± 2.8	12.1 ± 1.9	11.0 ± 1.6	13.5 ± 2.5
6	M	47	16.6 ± 1.4	16.5 ± 2.3	11.3 ± 2.1	12.2 ± 2.0	10.7 ± 2.0	12.4 ± 2.4
7	F	52	16.8 ± 3.0	18.0 ± 0.7	12.7 ± 3.4	12.4 ± 1.3	12.7 ± 2.2	12.9 ± 0.8
8	M	41	17.8 ± 3.3	18.4 ± 4.1	12.7 ± 2.9	15.2 ± 3.6	12.8 ± 1.9	16.1 ± 3.3
9	M	52	14.4 ± 1.7	14.3 ± 1.9	12.0 ± 1.0	13.0 ± 1.8	12.5 ± 2.0	13.3 ± 1.2
10	M	40	16.6 ± 2.4	17.6 ± 1.7	12.7 ± 2.3	15.0 ± 1.0	11.2 ± 1.8	14.9 ± 1.1
Mean			16.5 ± 1.4	17.1 ± 1.1	12.5 ± 1.6	13.6 ± 1.8	11.9 ± 1.3	13.5 ± 1.4

To visualize the differences in performance measurements between the two measurement techniques (stylus and Conoprobe), the Bland–Altman analysis was used and is shown in Fig. 6. The red region around the mean specifies the 95% confidence interval in the mean difference. The average difference between stylus data and Conoprobe data using the NRM was $-1.05\mathrm{mm}$ indicating the Conoprobe data performed worse than stylus data with NRM. The average difference between stylus data and Conoprobe data using NRM-VC was $-1.42\mathrm{mm}$ indicating the Conoprobe data performed worse than stylus data with NRM-VC as well. Moreover, using the same data, the different registration methods were compared and shown in Fig. 7. The average difference between the RRM and the NRM-VC using stylus data was 4.48 mm, indicating that the NRM-VC performed significantly better than the RRM when taking organ surface measurements with an optically tracked stylus. The average difference between the RRM and the NRM-VC using Conoprobe data was 3.66 mm indicating that NRM-VC performed better than RRM with Conoprobe data as well.

Fig. 6.

Bland–Altman analysis is used to visualize the differences in measurements between stylus and Conoprobe applied by (a) NRM and (b) NRM with vena cava feature.

Fig. 7.

Bland–Altman analysis is used to visualize the differences in measurements between RRM and NRM with vena cava feature applied on (a) stylus data and (b) Conoprobe data.

To better understand how the result is affected by target location, an analysis between the distance from target to surface data and TRE is applied and shown for each digitization method in Fig. 8.

Fig. 8.

Target location is analyzed to find the relationship between the TRE and the depth of the target. The blue shows the results of RRM. The orange shows the results of NRM. The gray shows the results of NRM with vena cava feature. The relationship between the TRE and the depth of the target is based on (a) stylus data and (b) Conoprobe data.

4. Discussion

In a previous study,⁷ Conoprobe-enabled scanning was found to be useful in the context of image-guided surgery under controlled phantom and ex vivo benchtop experiments. In this study, the calibration and reattachment experiment demonstrated that the attachment design was functional in the OR and stable with repeated use. The results presented represent the first effort to evaluate the potential of non-contact digitization method coupled to nonrigid registration to map liver deformation during open surgery using clinically acquired subsurface anatomical targets. With respect to subsurface anatomical target error, previous work by Clements et al.¹² reported on this and proposed a validation technique similar to the results herein. However, the work presented here is different in four important ways. First, the deformation correction algorithm employed here is LIBR by Heiselman et al.,⁴ which incorporates subsurface anatomical features (in this case, vena cava) in the registration process. Second, the previous study compared the 2D segmented target contour to 3D target models using mean closest point distances, which could underestimate true TREs due to the lack of correspondence between the two subjects. In this study, we proposed a centerline approach to ensure the correspondence and comparison are of good visual and quantitative fidelity. Third, this study represents an expanded patient cohort with an increased number of in vivo targets for analysis. It should be noted however that while the patient cohort has increased in size enabling more evaluation, the sample size is still quite limited and is not of sufficient size to study population effects. For example, parity on gender would be a natural first population study variant.¹⁷ Finally, the error induced by ultrasound probe pressure was considered and compensated for in this study. The improvements to methods and the analysis performed are important developments from the viewpoint of guidance system development, and it also emphasizes that intraoperative validation remains a challenging problem within the field. Although image-to-image alignment tasks allow for detailed comparisons with extensive quantitative metrics, image-to-physical error assessment within the in vivo clinical environment remains one of the most difficult barriers to overcome in the development of guidance technologies in procedural medicine. The work presented here is unique and novel in addressing these challenges.

With respect to the digitization methods themselves, the surface coverage obtained using the optically tracked stylus data was greater than that acquired from the Conoprobe by $\sim 5\%$. This correlates with previous benchtop experiments demonstrating the challenge of utilizing the Conoprobe to overcome line-of-sight constraints associated with the straight-path measurement configuration.⁸ The work here suggests that the relatively small degradation in coverage may effectively offset any potential benefit that the noncontact method might afford. The optically tracked stylus allowed for better coverage by enabling a digitization reach over the liver’s superior dome routinely. The versatility of being able to angle the stylus coupled to a longer shaft enabled more coverage. Interestingly, for patient 8, the surface coverage of Conoprobe is greater than that of stylus because, in this case, the surgeon could not measure under right rib cage due to alignment of camera with the cavity and orientation of the tracking spheres on each measurement device resulting in poor surface coverage. In this case, the design of the stylus shaft created an additional geometric constraint. It is important to note in this case however that despite the more considerable coverage by the Conoprobe, it did not outperform the stylus-based acquisition. This likely indicates additional factors that influenced our experience. For example, data quality with the Conoprobe has an SNR dependency, which is simply not present for stylus data. Although our past work provides us some experience in ensuring fidelity of Conoprobe measurements, the in vivo environment is much more dynamic than the bench and it is difficult to assess any effects that might be a result of the surgical theatre.

Figure 5 shows the qualitative visualizations representing the impact of the NRM with the vena cava feature on the alignment of the tracked iUS with the corresponding anatomical feature generated from the preoperative liver model. Patients 2 and 3 represent the best and worst targets in the study, respectively. The best target results are based on stylus data matches and over the course of the data in Table 1, and the results in Table 2, it suggests that the stylus may remain a more reliable localization device and produce better results. Patient 6 shows the best target alignment result based on Conoprobe data, which is not as good alignment as the best result based on stylus data. Interestingly however, patient 3 illustrates the overall worst alignment and was also based on the tracked stylus. In postprocedure analysis of the case, it was found that the surgeon had inadvertently placed their finger under the left lobe when measuring surface points with a stylus and did not do so when measuring surface points with Conoprobe.

To visualize the differences in measurements between two different instruments (in this case, stylus and Conoprobe), the Bland–Altman analysis was used in Fig. 6. The results indicate that the validation results by Conoprobe data were modestly worse than that of stylus data. Since the line of equality, in which the difference is equal to zero, is not within the confidence interval of the mean difference, the difference is significant. However, only clinical goals could define whether the agreement interval is too wide or sufficiently narrow for our purpose.¹⁸ Any difference smaller than the localization error, whether statistically significant or not, would likely not be practically significant. The mean difference under every situation is smaller than the calibration TRE of Conoprobe. Therefore, it is too early to rule out the potential of Conoprobe as a useful tool for surface data digitization. Moreover, previous work demonstrated that surface extent can affect registration results,⁸ and although the work here does follow a similar trend, it should be noted that previous work looking at data extent had much less variability in TRE as data extents became larger than 22% (see Sec. 2.1.2). Past this threshold, which reflects the average data extent for stylus and some cases for Conoprobe here, an asymptotic behavior emerged in that previous work.⁸ However, looking across Table 1, there is clear variability in extent, and some cases operated below this 22% threshold. In addition, we acknowledge that there are many factors associated with in vivo validation that could have influenced these results as well. Without modifying the Conoprobe to capture greater data extent on the organ surface and understanding SNR of the Conoprobe in the surgical theater, a definitive answer will be difficult to resolve. However, given the previous work that showed the benefit of non-contact methods of digitization, it does suggest that the process of surgery itself influences the comparison and in this case negated the positive aspects with noncontact digitization. Although the results in Table 2 demonstrates an average improvement of target error, the Bland–Altman analysis in Fig. 7 demonstrates the robustness of the correction process itself is not dependent on the surface digitization method. However, a strong bias is demonstrated to favor the deformable correction approach over conventional RRMs.

Finally, in addition to reporting the subsurface centerline-to-centerline distance metrics in Table 2, the relationships between the TRE and the distance from target to surface data have also been reported in Fig. 8 for better understanding. Similar to Fig. 7, Figs. 8(a) and 8(b) implies that nonrigid registration consistently outperforms rigid registration. The results in clinical data also demonstrate that targets close to physical data typically have reduced target error, which corroborates previous work by Heiselman, Jarnagin, and Miga⁴ that demonstrated a similar trend in a semisimulated phantom-to-human validation dataset. The improvements associated with including the vena cava feature, i.e., a single slice of ultrasound data, in the NRM, are quite subtle and do not rise to the level of significance in this cohort. It is likely additional subject numbers may approach significance.

5. Conclusion

This study demonstrates the functionality of the Conoprobe attachment design in the OR and evaluates the feasibility of the Conoprobe as a noncontact digitization method in open liver surgery. The surface coverage of Conoprobe was consistently less than that of the stylus. Although the work here establishes the systematic improvement in registration in human systems using our nonrigid methods over standard-of-care rigid surface registration methods, it also makes clear that methods of intraoperative digitization still need to improve and develop in order to take advantage of the benefits of noncontact digitization methods performed at the bench.

Biographies

Bowen Xiang is a fourth-year biomedical engineering PhD candidate at the Vanderbilt University. He received his bachelor’s degree in biomedical engineering from the University of Virginia. His current research is focused on tumor localization for liver surgery.

Biographies of the other authors are not available.

Disclosures

No conflicts of interest, financial, or otherwise, are declared by the authors.

Code and Data Availability

Some preliminary data have been reported in Ref. 19. Data used in this paper were acquired in accordance with an IRB approved study and its availability is compliant with NIH policy and the protection of patient data policies.