A Novel Flexible Endomicroscopic Scanning Instrument with Visual Servoing Control

Optical biopsy using probe‐based confocal laser endomicroscopy (pCLE) and optical coherence tomography (OCT) technologies performs in vivo tissue scanning of organ surfaces and plays a crucial role in early cancer diagnosis. Herein, a flexible endomicroscopic scanning instrument based on a 3‐degree‐of‐freedom flexible hinge structure and visual servoing control is presented. The instrument integrates OCT and pCLE. OCT information compensates for the imaging distance of pCLE, which can combine the advantages of the two technologies. The mechanism allows for tissue surface scanning with a wire‐driven active bending part and dynamic adjustment of the probe‐tissue distance using a combination of wire and spring‐driven, which satisfy large‐area scanning while maintaining consistence tissue contacts. Local optimization‐based visual servoing control is proposed to optimize the tissue scanning. Through the scanning experiments, high‐quality mosaics of uneven surfaces are obtained with an effective area that is larger than 4 mm2. In vivo animal trial further verifies the feasibility and clinical potential of the instrument.


Introduction
Gastrointestinal cancers (GI) seriously threaten human health and life without distinctive early symptoms and are frequently discovered at a late stage. [1]However, early diagnosis of cancer can greatly improve the survival rate by over 90%. [2]athologic diagnosis is still the primary means of diagnosing early cancer, but the surgical process of tissue acquisition brings many risks such as bleeding and incorrect tissue sampling. [3]ptical biopsy based on probe-based confocal laser endomicroscopy (pCLE) can detect tissue in situ and in real time and become a promising technology for early cancer diagnosis. [4]ue to the advantage of real-time imaging, pCLE has been applied in clinical diagnoses such as GI tract, [5,6] thyroid gland, [7] and lung. [8]he technology of optical coherence tomography (OCT) allows cross-sectional imaging and provides depth information on tissue structures.It has been applied in vitreoretinal surgery, [9] bladder cancer diagnosis, [10] and neurointerventional surgery. [11]OCT also can be used to guide the movement of surgical devices such as detecting the depth of needle insertions [12] and guiding surgical instruments to complete optical biopsy. [13]owever, due to its narrow field of view (FOV, approximately 200-600 μm), [14] it is necessary to collect large area microscopic images for image stitching. [15]It is generally believed that the acquisition of large-scale microscopic images requires large-area microscopic tissue scanning with a robotic instrument. [4,16,17]To obtain large-scale image, Erden et al. scanned spiral and raster trajectories of ex vivo bovine liver tissue and chicken breast tissue using a six-axis industrial robot (TX40, Stüaubli, Faverges, France). [18]rden et al. also achieved soft tissue scanning with TX40 robot to obtain a large-scale pCLE image. [19]Giataganas et al. used a KUKA robotic arm (KUKA Roboter GmbH, Augsburg, Germany) to perform surface scanning automatically and acquired consistent tissue images. [20]Although industrial robots have achieved large-area scanning, the size limitation of instruments constrains their application in clinical fields.To achieve better clinical application of pCLE, researchers have studied delicate and miniaturized robots with proper control for minimally invasive surgery.
Giataganas et al. proposed a handheld automated scanner that uses a cam mechanism to drive the high-resolution imaging probe and achieve efficient and large-area scanning of tissues. [21]uo et al. developed a gear-guided linkage scanning mechanism to achieve large-area scanning of the breast cavity. [22,23]Although researchers have studied miniaturized surgical robots, these instruments lack suitable control to achieve precise scanning.Rosa et al. proposed a visual servoing control for tissue scanning. [24]The experimental verification of visual servoing control on the TX40 robot demonstrated the optimization for consistent tissue scanning.To enhance the robustness of the control, Rosa et al. developed an optimization control based on Kalman filter, allowing a reliable registration even under nonsmooth sample scanning process. [25]Xu et al. also proposed a position-based visual servoing control to conduct circular and spiral scanning on ex vivo porcine stomach tissue. [26]However, these methods only achieved optimal control in the direction of tissue surface without consideration of the perpendicular direction to the surface.To achieve large-area mosaicking, consistent and stable tissue contact is very important. [19]An optimal control is required to maintain a steady contact between the probe and tissue.Zhang et al. proposed a contact control based on the depth information of OCT image. [27]However, this control performed poorly when the velocities of tissue were faster than 30 μm s À1 .
Most importantly, the expected instruments should achieve in vivo tissue scanning through the GI tract.The scanner combined with industrial robots [24,25] or Da Vinci [27] and a linkage-driven mechanism [22,23] are difficult to adapt to the flexible and curved structure of the GI tract.Continuum robots have been widely used in the field of medical surgery, becoming one hope to scan in vivo tissue through the GI tract. [28,29]Conventional continuum is mainly based on the rolling contact joints, [30] spherical joints, [28] and pin joints. [31]Compared with the traditional articulated continuous mechanism, the compliant mechanism has obvious advantages such as a skeleton and convenient assembly. [29]Hence, the driving methods of a continuum instrument usually include built-in motors, pneumatics, and wire-driven.The wire-driven is a preferred choice for medical usage due to its simple structure and minimization.
From the earlier considerations, we developed a flexible scanning instrument based on a compliant mechanism and wire-driven to conduct in vivo tissue scanning.This instrument combines two probes (pCLE and OCT) to meet the requirements of three-degree-of-freedom (DOF) motion (see Figure 1).During the scanning, a design of combining spring and driven wire achieved linear movement of the probe and adaptively adjusted the contact distance.The local optimization-based visual servoing control method is adapted to the flexible instrument, and the effectiveness has been verified by the phantom scanning experiment.Moreover, in vivo experiment demonstrates the clinical potential of the instrument.

Experimental Section
The flexible endomicroscopic instrument is designed to scan a large area of tissue surface (typically more than 3 mm 2 ) [18] and achieve real-time optical biopsy through natural orifice transluminal endoscopic surgery.A long flexible passive bending part is needed to transmit the driven force to the active bending part.Based on the earlier considerations, a flexible endomicroscopic scanning system with the adaptive visual servoing control method is developed in this section.The whole scanning system consists of the control system, optical system, and flexible scanning instrument, as shown in Figure 2.

Design of Flexible Instrument
The instrument consists of the actuation and execution units.An active bending part coupled with an end-effector and flexible guide shaft comprise the execution unit.The overall length of the instrument (executive unit and actuation unit) is 1100 mm, satisfying the tissue scanning through natural orifice transluminal endoscopic surgery.The porotype of the instrument with an enlarged view of the active bending part is shown in Figure 3a.According to the target trajectory scanned by the flexible instrument, the direction of bending is defined as yaw (Y) and pitch (P).This work chooses the arrangement sequence of YPYPYPYP (see Figure 3b) that can realize the bending in four directions on two axes, meeting the motion requirements on tissue surfaces.
The flexible hinge structure is the main part of the active bending part.The internal working channel of this structure needs to accommodate two imaging probes, OCT and pCLE.We design flexible hinge structures based on the super-elastic properties of 55-Nitinol to realize bending motion at room temperature.The inner diameter of the instrument is designed as 6 mm based on the probe size.The wall thickness of the active bending part is set to 0.25 mm, resulting in an outer diameter of 6.5 mm.The maximum bending angle of the flexible hinge unit is set to 10°, and eight flexible hinge units are evenly arranged in the yaw and pitch directions.Based on the assumption of a constant curvature model, the entire instrument can achieve a 40°deflection angle, meeting the requirements of the scanning area (greater than 3 mm 2 ).The main parameters of the flexible hinge structure are shown in Table 1.
The finite element analysis of the active bending part is performed in the workbench platform of ANSYS software to evaluate the maximum strain of the compliant mechanism and the material properties are shown in Table 2.In the simulation, the entire active bending part is divided into tetrahedral mesh units with a total of 19 438 elements.Since the main deformation area is occurred in a flexible hinge beam, a more detailed mesh division is performed at the beam.A driven force that can cause a maximum deflection angle is applied along the direction of the driven wire to simulate the torque provided by the driven wire.When the flexible hinge unit beam is bent to a maximum deflection angle of 10°, the maximum strain is 2.9%, as shown in Figure 3c, which meets the maximum strain (12%) requirement of nitinol alloy.
The designed instrument accommodates two imaging probes installed side by side.The end-effector is fixed with soft rubber and positioning pins for the OCT and pCLE probes.The rubber is placed in a connecting tube, and another end of the connecting tube connects to a spring.Hence, the spring is connected to a retaining ring, and the retaining ring is fixed on the flexible hinge through positioning pins.Two rows of positioning pins on the connecting tube, cooperating with the sliding slot on the active bending part, realize the directional limitation of the probe's axial linear displacement.The spring is compressed by the tension force of the driven wire; while the force of the wire is unloading, the spring returns to its original position.Linear displacement is used to adjust the contact distance between the probe and the tissue during the scanning process.The design of the endeffector is shown in Figure 4.
This work chooses a multilumen tube as the flexible guiding shaft of the flexible instrument.The multilumen tube is made of low-density polyethylene plastics, 13 lumens customized by grinding tools according to the size of the driven wire.The outer diameter of the multilumen tube is set to 7.6 mm, and the length is 1000 mm.There are 12 channels of the driven wires with a diameter of 0.3 mm distributed uniformly in a ring on the cross-section of the tube wall (see enlarged view in Figure 4).The center lumen is used for passing through the optical probes.At the end of tube, a groove with a diameter of 6.5 mm and a depth of 1 mm is processed to achieve a stable connection and fixation with the active bending part.The actuation unit consists of four modules: supporting plate, winding wheel for driven wire, tensioning wheel, and motors.As shown in Figure 5a, a motor is fixed to a spiral winding wheel, and two sets of driven wires are wound in opposite directions on each wheel.When the motor is rotating, one set of driven wires tightens, while the other set relaxes, achieving one-sided bending force on the instrument.Thus, two motors control the xand yaxis motions without interfering with each other.The third motor controls the z-axis motion.The rotation of the wheel tightens the driven wire and compresses the spring.When the winding group rotates in the opposite direction, the driven wire relaxes and the spring returns to original position, realizing the linear displacement in the z-axis direction.The arrangement of the driven wires and winding wheels is shown in Figure 5b.

Optical System
The optical system includes pCLE and OCT imaging units.In terms of pCLE imaging unit, the diameter of the pCLE probe (Mauna Kea Technologies, France) is 1.4 mm with an FOV of 600 μm.The probe is coupled to a confocal endomicroscopy of an in-house design and is based around a commercial laser confocal imaging system (CLS, Thorlabs USA). [23]he OCT imaging unit uses a custom probe-type imaging fiber (diameter 3.3 mm), which can achieve endoscopic scanning. [32]The end of the probe is connected to the OCT imaging system (OCT Viewer, Light Vision Technology, China).OCT and pCLE probes are fixed on the instrument side by side, as shown in Figure 3a.When installing two optical probes after the pCLE probe is fixed, the OCT probe is rotated to change the imaging area of OCT.The imaging areas of the two probes must be adjusted until the OCT's imaging area (scanning line) passes vertically through the pCLE's imaging area, ensuring that they image the same region.

Control System
The control system coordinates with pCLE imaging control, OCT imaging control, and motor motion control.The main function of the control system is to ensure real-time display of images from the pCLE imaging system and OCT imaging system, as well as online real-time processing of collected image data to obtain control parameters for motor input.
The program of the scanning control employs a multithreaded running mode to achieve the coordinated completion of the scanning process among various functional units.Each subprogram runs asynchronously in its own threads, and cooperation between threads is achieved through necessary data communication.
pCLE and OCT imaging control threads transmit image data by a queue structure and complete real-time image display.The acquisition frequency of the imaging units (pCLE and OCT) is 8 fps, which satisfies the requirement of image data to analyze motion control information.The main thread reads the image data from the queue for processing and analyzing the motion control parameters of the motor.Multithreaded mode ensures the smoothness of the implementation of the control system  and avoids confusion caused by mutual interference between various functions.

Visual Servoing Control Model
Various factors constrain the control accuracy of flexible instruments, and traditional control methods make it difficult to achieve the motion accuracy required for tissue scanning.Therefore, it is necessary to detect the actual position of the instrument in the image coordinate system to optimize its movement trajectory of instruments.Based on the actual instrument position information, we proposed the local optimization-based visual servoing control to achieve the precise control of the instrument.An iterative process for optimizing the probe trajectory has been added to achieve better mosaicking results.At the same time, to reduce the impact of accumulated errors, the local optimization method is used to calculate feedback parameters.
Local optimization includes the optimization in the xÀy plane and the axial direction (z-axis) of the instrument.In the direction of z-axis, OCT-based visual servoing control is used to maintain the steady contact between the probe and tissue.The probe moves to the optimum imaging position of pCLE probe, according to the command of position input Δt that is calculated by Equation ( 1): where k is the transformation coefficient of the controller from pixel in OCT image to actual distance; D is the optimum imaging position of pCLE probe that can be measured by testing the imaging performance at different positions; and d is the current contact distance between the probe and tissue.
To obtain the contact information, the grayscale value (GðlÞ, average value of pixels in row l) of the OCT image corresponding to the target area under the pCLE probe is obtained.The maximum g max and minimum mean square errors g err of Gð:Þ are obtained by Equation ( 2) and (4): ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi ðGðiÞ À GðlÞÞ 2 q , i ¼ 1, 2, : : : , M À 1 (3) The current contact distance d is first value greater than threshold thr, calculated by Equation ( 5) and ( 6): d ¼ minðjÞ, where GðjÞ > thr (6) In the xÀy plane, the distance between the current point and the next point is calculated by Equation ( 7) and taken as the optimization goal: where P Ã ð:Þ is predefined scanning trajectory points.X i and Y i are row coordinates and column coordinates, respectively.The initial bending motion of the probe is controlled based on the initial motion parameters t.After the bending motion is completed, z-axis optimization is carried out to optimize image distance, and pCLE images are acquired until the contact distance optimization is finished.Then, the displacement of the image S is calculated by the normalized cross correlation-enhanced (NCC-E) in ref. [26] to obtain the actual deflection distance of the probe.Based on the actual deflection distance, the parameters for the next step are calculated according to Equation (10): After obtaining the new control parameters, it is determined whether they are less than the minimum adjustable range.We found experimentally that the error of imaging distance less than 10 μm from optimal value was necessary to ensure clear enough image results.After the parameter of t is less than 10 μm, the optimization goal is achieved.The optimization of the next position can be carried out when the optimization of current position is achieved.If not, the optimization of this point continues based on the new control parameters.The flowchart is shown in Figure 6.
The flexible instrument implements a bending motion through the driven wire.Due to the lack of rotational motion, it is easy to achieve the scanning process of the raster trajectory.In the process of raster trajectory tissue scanning, the bending motion of the instrument alternates between two axes.After one axis bends a certain distance, the other axis adjusts the spacing between the two rows of scanning trajectory.When using local optimization control for scanning, only one axis movement is required for the position change of each point in the trajectory, which reduces the mutual influence of two axes moving simultaneously.

Mechanical Performance
Statistical methods were used to determine the relationship between the driven wire tension and the deflection angle of the instrument based on experimental data.The experimental setup is shown in Figure 7.The actual deflection angle was measured by a high-resolution digital camera setup over the tip.
Wire tension was measured in real-time using a six-axis force sensor (Nano, resolution of 3.125 mN, ATI Industrial Automation Company, USA) that was fixed on a linear moving stage.Six trials were conducted, each consisting of experiments of pitch and yaw bending performance.The relationship between the driven wire tension and the deflection angle of the instrument was obtained from the average of the experimental data, as shown in Figure 8.
For pitch bending motion, when the instrument reached the same deflection angle, the maximum deviation of the wire tension between moving forth and back was 0.28 N, and the minimum deviation of that was 0.04 N.For yaw bending motion, when the instrument reached the same deflection angle, the maximum and minimum deviations of the wire tension between moving forth and back were 0.19 and 0.06 N, respectively.These indicate a good consistency between the driven wire tension and deflection angle.
When the flexible instrument bent to its maximum angle of 40°, the force of the wire was 1.04 N. Wire force was much lower than the breaking strength of the seven-strand steel wire, indicating that it meets the technical requirements.
For assessment of the bending performance, measurements were made in two parts: (a: move forth) bending from 0°to 30°and (b: move back) returning to 0°.The maximum hysteresis of the deflection angle is 2.8°between moving forth and back the same effective length of the driven wire in the pitch direction, and the minimum hysteresis of deflection angle is 0.2°.In the yaw direction, the maximum hysteresis of the deflection angle is 2.67°, and the minimum value of the deflection angle is 0.3°.Repeatability measurements were performed over six trials, measuring the deflection angle of pitch and yaw directions.The average and a third-order exponential fitting were made from experimental data, obtaining the kinematic relationship, as shown in Figure 9.The repeatability and positioning accuracy of the instrument are presented in Table 3.

Scanning Parameter Calibration
The difference between the actual scanning trajectory and the expected trajectory is large during the scanning process of the flexible instrument.It is difficult to achieve high motion accuracy required for tissue scanning based on the parameters solved by the kinematics.Motion parameters can be analyzed by statistical method from multiple trials.Calibration experiments of motion parameters were performed to find the motor input parameters matching with the displacements of the image.The experimental setup is shown in Figure 10, the phantom of sponge was scanned.
The motor was controlled to drive the bending motion of the instrument under the position control mode.After each completion of the same distance for motor moving, the pCLE and OCT images are captured.For each movement, the input of the motor is 2000 units.After a series of movement, pCLE and OCT image sequences were obtained.The xand y-axis motor movements correspond to the yaw and pitch motion, respectively.The z-axis motor movement corresponds to the axial motion.
The displacement of the pCLE images obtained by NCC-E is shown in Figure 11a,b.Figure 11c reveals imaging distance in OCT images calculated by Equation (6).Although the tensioning state of the driven wire is affected, resulting in discontinuous motion in Figure 11b.There is a linear relationship between the image displacement and motor movement, as shown in Figure 11.Thus, an approximate formula can be derived by fitting the measured data, as shown in Equation ( 11).Dis im ¼ a Ã x mo (11)   where Dis im is image displacement, x mo is movement of motor, and a is the transformation coefficient that scales the input of the motor to the displacement of image in pixel.In the pitch and yaw directions, the transformation coefficients are 0.019.In the axial direction, the transformation coefficient is 0.0008.Fitting errors and fitting degrees reflect the error of mathematical model, as presented in Table 4. Higher fitting degrees represent the higher correlation between motor movement and image displacement.

Phantom Experiments
Before the scanning experiment of visual servoing optimization control, the scanning experiment of open loop control was carried out.A sponge structure phantom was scanned with a raster trajectory (setup in Figure 10).Experimental results of open loop control obtained a tiny effective area of mosaic.The image trajectory was irregular and did not match the target trajectory, as shown in Figure 12a.This indicated that the open-loop control cannot effectively achieve the tissue scanning.
To verify the feasibility of the local optimization-based visual servoing control that proposed in this work, a comparative experiment with global optimization-based servoing control was conducted, and collected an image sequence.In this trial, two axes of the instrument were optimized simultaneously at each movement, causing the significant coupling effects of the two axes on motion control.This control reduces the positioning accuracy of the instrument.Figure 12b shows the scanning results of global optimization-based visual servoing control.Scanning results include the mosaicking result from the acquired image sequence and the image trajectory of the mosaicking process.
From the scanning results of global optimization with two axes, it can be concluded that the image trajectory is chaotic when two axes are optimized simultaneously.The movement of one   axis affects the positioning accuracy of the other axis, thereby affecting the overall positioning accuracy of the instrument.
To reduce the simultaneous movement of the two axes affects the positioning accuracy of the instrument, scanning experiments of single axial optimization were performed.During the scanning, motion instructions were only transmitted to the active axis that needed to move.Thus, the single-axis motion for each movement is achieved.Figure 12c shows the image trajectory and mosaicking result.
The image trajectory of the single axial optimization is relatively smoother compared with the optimization with two axes because there is no interference on the active axis from the other axis.However, under the global optimization control, the cumulative error causes reverse motion at the junction of two rows of trajectory.Because the x-axis motion causes the displacement in the y-axis direction, resulting in reverse motion when the instrument bends in the y-axis direction.However, the reverse motion   in the y-axis affects the tensioning state of the driven wire and contributes to difficulties for subsequent scanning.
Considering the problem of reverse motion caused by cumulative errors, the local optimization-based visual servoing control was used in scanning experiments.In this trial, the transmitting function of motion instructions for the nonmoving axis was turned off.The scanning results are shown in the Figure 12d.The image trajectory of the local optimization-based visual servoing control is closer to the raster scanning trajectory compared with the global optimization-based visual servoing control.The impact of reverse motion is avoided, ensuring the consistency of the driven wire tensioning state during the scanning process.The trajectory errors of scanning are shown in Table 5, and local optimization with a single axis reaches the minimum trajectory error among the four control methods.It can be found from the experimental results that the effective area of local optimization is the largest and reaches 4.36 mm 2 .
In the z-axis direction, four trials (open loop control, global optimization with two axes, global optimization with single axis, and local optimization with single axis) all adopt the visual servoing control method based on OCT depth to achieve steady contact between the probe and the tissue.In the four groups of scanning experiments, the surface positions in OCT images are shown in Figure 12.The contact distance is maintained within four pixels (40 μm) during the scanning process.A quantitative analysis of contact distance errors is shown in Table 5. Utilizing OCT-based visual servoing control, the adjustment of the distance between the probe and tissue can be executed in real time with high accuracy.It ensures uninterrupted and continuous scanning of uneven surfaces.

In Vivo Animal Experiments
Scanning experiments were performed to verify the feasibility of the instrument in porcine model with a weight of 50 kg.This study was reviewed and approved by the Ethics Committee of Teda International Cardiovascular Hospital (Approval: TICH-JY-202205027-4).The experimental animal was under general anesthesia.The experimental setup is shown in Figure 13a.
The flexible instrument integrated a Cellvizio device (Mauna Kea Technologies, Paris, France) and OCT system (OCT Viewer, Light vision technology, China) was used to scan the acriflavine-stained rectum tissues.
The scanning trajectory of pCLE was set as a raster trajectory.The pCLE images were acquired and generated a large-scale mosaic.The target area was also scanned by OCT, with a scanning range set to 3 mm Â 3 mm.The OCT image sequence was  rendered by the library of visualization toolkit.Different transparencies were set corresponding to the different grayscale values of the pixels in the OCT image.According to the actual physical size of target area, the fusion results combine the obtained OCT volume rendering and pCLE mosaic.We performed the raster scans six trials, and experimental data are listed in Table 6.Two pCLE mosaics with a large number of back-toback, ring-shaped, and symmetric twin crypts of similar diameter cells with hyperfluorescent borders depict the typical appearances of colonic cells in Figure 13b.OCT volume rendering result with a cross-sectional view is shown in Figure 13c.The fusion result displays the integration of the two endomicroscopic imaging modes in Figure 13d.

Discussion
This work developed a novel flexible instrument by combining capability of three-DOF motion and accommodating pCLE and OCT to achieve large-area tissue scanning.Tissue scanning of pCLE with a flexible endomicroscopic instrument is a great challenge due to various constraints, such as assembly errors, system errors, tissue contact force, and changes in driven wire tensioning state.Optimizing control methods plays a key role in completing tissue scanning that meets the target trajectory.This work uses a visual servoing control to improve scanning results and obtain large-scale pLCE mosaics.Raster trajectory is used for scanning experiments to verify the feasibility of the proposed control method.The dimensions of the instrument are smaller than most commercial endoscopes (the outer diameter is about 10 mm).In clinical practice, the instrument could also reach the target through the working channel of the endoscopic platform, such as direct drive endoscopic system (DDES, developed by Boston Scientific, MA), [33] rigid and flexible outer sheaths, [34,35] the incisionless operating platform (IOP, developed by USGI Medical, CA), [36,37] or STIFF-FLOP flexible manipulator. [38]These platforms are equipped with working channels that are more than 6.5 mm in diameter.The outer diameter of the instrument is designed according to the size of probes and driven wires.In the future, smaller size probes will be researched to increase the flexibility and reduce the outer diameter of the instrument.
Evaluation of mechanical performance is conducted to verify the feasibility of the instrument.The repeatability and positioning accuracy of the instrument reaches 0.28°for the pitch direction bending motion and 0.37°for the yaw direction bending motion on average.High accuracy is achieved with 0.097 mm for pitch direction bending motion and 0.129 mm for yaw direction bending motion at the tip, respectively.The FOV of commercial pCLE probes is between 0.24 and 1 mm in diameter.It is acceptable accuracy (0.129 mm) for achieving nongaps mosaicking of the raster trajectory and should be satisfactory for diagnosis purposes.Moreover, the feasibility of visual servoing control of the instrument was also verified through experimental evaluation.
The current handheld rigid endomicroscopic scanning devices [21][22][23] achieved a tip positioning accuracy in the range of 0.2-0.25 mm.The flexible endomicroscopic scanning device [29] achieved a tip positioning accuracy of approximately 0.33 mm.Compared with these devices, the proposed flexible instrument achieved a higher accuracy (0.129 mm).The reasons for high tip positioning accuracy in this work can be considered as the following.The small beam height ratio reduced the maximum deflection angle (10°) of each joint and improved the accuracy of bending.Due to the integrated processing by high-precision laser cutting, the active bending part of instrument required minimal assembly and few additional components.A small assembly error combined with a high-precision machining method also improved the motion accuracy.
To achieve precise control of the flexible instrument, the local optimization-based visual servoing control is proposed.During the scanning, accumulative errors have a significant impact on scanning results and lead to reverse motion near the target position.Reverse motion poses a great challenge to the tensioning state of the driven wire, which seriously affects the control accuracy of the instrument.Therefore, to prevent reverse motion, a visual servoing control method based on local optimization is used.The global optimization goal is decomposed into local optimization goals.Each optimization goal is a local correction based on the current position.This method reduces the deviation caused by the reverse motion of the global coordinates.
The phantom scanning experiments using the proposed instrument verify the performance of visual servoing control based on local optimization.Only one axis bending is needed to move to the next point in the raster trajectory, which avoids the motion coupling caused by the simultaneous movement of two axes.The movement of the nonmoving axis is restricted.Experimental results show that one axis motion optimizes the image trajectory.Large-scale mosaic requires a total area of typically 3 mm 2 , which is within a circle with 1 mm diameter. [18]The effective area of local optimization-based visual servoing has reached 4.36 mm 2 and meets the area requirements of clinical diagnosis.Compared with frozen section (40 min) and intraoperative specimen X-rays (10 min, but it has a low accuracy), the proposed method takes approximately 16 min to scan the target surface, which is efficient and meets the needs of clinical application.
OCT plays a crucial role in adjusting the contact distance between the probe and tissue.During the entire scanning process, the contact distance maintaining steady and high-quality pCLE image sequences are acquired.Large-scale mosaics are successfully generated from the pCLE image sequences.Without the OCT visual servoing control, the instrument movement tends to move him away from the tissue.And the large distance between the probe and tissue makes it impossible to obtain clear images.OCT produces a lateral resolution of approximately 10 μm.
Compared to other methods such as ultrasound (lateral resolution approximately 50 μm), [39] and laser displacement sensors (difficult to minimize the size under 6.5 mm in diameter), [40] it satisfies the precision of distance detection.pCLE and OCT images are simultaneously acquired, which improves the efficiency of surgical procedures.
In vivo animal experiments were performed to verify the largearea scanning and imaging with the pCLE and OCT probes.The instrument that integrated the optical system scanned the tissues of in vivo porcine rectum.pCLE image sequence and OCT image sequence were obtained through the scanning process.The fusion results combine the OCT volume rendering and pCLE mosaic.
This work presented a novel flexible endomicroscopic scanning instrument with visual servoing control to achieve a wide image scanning for optical biopsy surgery.[43] The proposed instrument is to address the problems associated with mechanical scanning over large tissue areas in confocal endomicroscopy, which can greatly improve the prospects for intraoperative in vivo digestive tract margin evaluation.

Conclusion
In this study, we have developed a flexible instrument to obtain continuous and high-quality pCLE image sequences using visual servoing control.To achieve high-precision of motion control, an integrated control system is developed and local optimizationbased visual servoing control is adopted.Phantom scanning experiments verify the effectiveness of the proposed control method.The large area scanning results (4.36 mm 2 ) without gap meet the early clinical diagnosis of GI cancer.The OCTbased visual servoing control maintains the steady contact between the probe and tissue, ensuring the clarity of each pCLE image.In vivo animal experiments demonstrate the feasibility and clinical potential of the instrument.

Figure 1 .
Figure 1.Concept of tissue scanning in the gastrointestinal tract used a flexible instrument.

Figure 2 .
Figure 2. The schematic diagram of the scanning system contains instruments, optical, and control systems.

Figure 3 .
Figure 3. Flexible instrument: a) prototype; b) active bending part; and c) finite element analysis.

Figure 4 .
Figure 4. Structure design of execution unit.

Figure 5 .
Figure 5. Actuation unit: a) structure design; b) arrangement of driven wire and spiral winding wheel.

Figure 6 .
Figure 6.The flowchart of the visual servoing control based on local optimization.

Figure 7 .
Figure 7. Experimental setup for testing mechanical performance.

Figure 8 .
Figure 8. Relationship between the driven wire tension and deflection angle: a) pitch direction; b) yaw direction.

Figure 9 .
Figure 9. Relationship between the driven wire displacement and deflection angle: a) measured value in pitch direction; b) fitting curve in pitch direction; and c) measured value in yaw direction; and d) fitting curve in yaw direction.

Figure 10 .
Figure 10.Experimental setup: a) actuation unit; b) end-effector; and c) image of the entire setup.

Figure 11 .
Figure 11.Relationship between the image displacement and movement of motor: a) x-axis; b) y-axis; and c) z-axis.

Figure 12 .
Figure 12.Scanning results: a) open-loop control; b) global optimization with two-axis control; c) global optimization with single-axis control; and d) local optimization with single-axis control.

Figure 13 .
Figure 13.In vivo animal experimental: a) setup; b) pCLE mosaics; c) OCT volume rendering result; and d) fusion result.

Table 1 .
The selected parameters of the flexible joint.
able 2. Material parameters of nitinol alloy.Parameter Value Young's modulus of austenite 70 GPa Poisson's ratio 0.33 Maximum residual strain 0.05 Forward strain start stress 342 MPa Forward strain end stress 443 MPa Reverse strain start stress 187 MPa Reverse strain end stress 120 MPa

Table 3 .
Result of scanning repeatability measurement.

Table 4 .
Fitting results of movement parameter.

Table 5 .
Trajectory errors of scanning.

Table 6 .
Tissue scanning data from animal experiments.