Large area robotically assisted optical coherence tomography (LARA-OCT)

We demonstrate large-area robotically assisted optical coherence tomography (LARA-OCT), utilizing a seven-degree-of-freedom robotic arm in conjunction with a 3.3 MHz swept-source OCT to raster scan samples of arbitrary shape. By combining multiple fields of view (FOV), LARA-OCT can probe a much larger area than conventional OCT. Also, nonplanar and curved surfaces like skin on arms and legs can be probed. The lenses in the LARA-OCT scanner with their normal FOV can have fewer aberrations and less complex optics compared to a single wide field design. This may be especially critical for high resolution scans. We directly use our fast MHz-OCT for tracking and stitching, making additional machine vision systems like cameras, positioning, tracking or navigation devices obsolete. This also eliminates the need for complex coordinate system registration between OCT and the machine vision system. We implemented a real time probe-to-surface control that maintains the probe alignment orthogonal to the sample by only using surface information from the OCT images. We present OCT data sets with volume sizes of 140 × 170 × 20 mm3, captured in 2.5 minutes.


Introduction
Optical coherence tomography (OCT) had a significant impact on medical imaging and nondestructive testing by enabling high-resolution cross-sectional imaging with millimeter depth penetration and micrometer axial and lateral resolutions [1,2].OCT operates on the principle of interferometry, which enables it to measure the echo time delay of backscattered or reflected light to construct images [1].It has been used extensively in ophthalmology [3,4], cardiology [5,6], dermatology [7,8], endoscopy [9], as well as other areas like non-destructive testing [10][11][12], process monitoring [13], and additive manufacturing [14], making it a valuable tool for non-invasive diagnostics and research.While conventional OCT has been proven valuable in examining small regions (< 10 × 10 mm 2 ) with exceptional detail, it may not always be ideal for capturing images of larger areas, such as whole organs, tissue sections, or large industrial samples.
However, in dermatology, inflammatory skin conditions often spread over large areas, such as psoriasis [15][16][17][18][19][20] or atopic dermatitis [21,22].It has been shown that OCT can accurately measure the epidermal thickness of psoriasis lesions, allowing for quantifiable imaging biomarkers that aid in making therapeutic decisions [17,19,23].Moreover, in non-destructive testing of non-biological samples such as wind turbine blades [11], marine coatings [12], monolithic data storages [24], or additive-manufactured parts [14], large-area imaging proves advantageous in detecting hidden defects or anomalies over the entire sample surface.Scanning multiple areas using hand-held or stationary OCT probes to capture these large regions would be time-consuming and laborious.The OCT system operator may concentrate on only a limited number of sparse surface locations to decrease the imaging procedure duration.However, this increases the possibility of missing relevant structures and pathologies.Furthermore, hand tremor can result in motion artifacts for hand-held OCT probes, especially with slow OCT engines at A-Scan rates of 1 MHz or below.Also, performing complex scan patterns by hand with a defined overlap for reliable image stitching might be almost impossible.
In previous studies, multiple techniques have been utilized to expand the lateral field of view (FOV) of OCT.These approaches are distinct from expanding the lateral FOV of a single OCT scan through the use of wide-angle camera lenses [25] or the acquisition of multiple single FOVs using translational motion stages [26][27][28] or robotic arms [24,[29][30][31][32][33][34][35][36][37] for creating a mosaic of individual scans in post.Song et al. demonstrated an OCT system that could scan an approximately 200 × 200 mm 2 area in one scan by using a wide-angle camera lens and a 100 kHz swept source laser with a 500 × 500 A-Scans resolution within ∼2.5 s [25].The slow sweep repetition rate allowed for a sufficiently large axial imaging range, enabling imaging of curved surfaces to an extent.However, the low A-Scan rate leads to unavoidable motion artifacts and requires sparse sampling for shorter acquisition times.While suitable for stationary samples, diagnosing biological tissues in-vivo presents challenges due to motion artifacts.Le et al. presented a hand-held OCT probe mounted on a translational and rotational stage used for large-area imaging of the dental region [28,38].An image is created by scanning six teeth with three individual OCT volume scans, which differ in viewing angle.The overall acquisition time for the mosaic is approximately 20 to 30 minutes using a swept source with a 200 kHz sweep rate.To minimize sample motion, a chin-and headrest in conjunction with a bite bar are used, which may cause some discomfort for the patient.Göb et al. introduced a translational stage with micrometer resolution for large-area imaging and vascular contrast imaging in combination with MHz-OCT systems [26,27].The device can capture a mosaiced FOV up to 200 × 200 mm 2 using single mosaic tiles of ∼12 × 12 mm 2 or with an interplay of stage and scanner movement resulting in elongated single scans.Processing delays currently constrain the theoretical maximum achievable imaging speed to less than 1 cm 2 /s.Since translational stages are limited to translational movements, imaging of an arbitrarily shaped surface is not feasible.Additionally, the sample must fit under the probe of the stage, which generally restricts the usability of translational stages.
To address these challenges, systems were developed that use robotic arms with six or seven degrees-of-freedom (DOF), with an OCT probe mounted as an end effector.For example, Huang et al. developed an OCT system based on a 7-DOF robotic arm with a 76 kHz A-Scan rate spectral domain OCT (SD-OCT) [34].They demonstrated the capability of this system to perform large-area imaging of planar and curved surfaces with a mosaiced FOV of up to 67.8 × 4.0 × 2.8 mm 3 composed of 18 individual volumes.The entire imaging process lasted for approximately 11 minutes and resulted in an imaging speed of 0.004 cm 2 /s.The path planning and positioning of the OCT probe were accomplished using a depth camera and an RGB camera that acquired point clouds before the measurement.An OCT volume was captured at each position to refine the positioning, and the robot pose was optimized.This process necessitates precise registration of the coordinate systems of each modality and can only be used on stationary samples.Ma et al. proposed a comparable design for monitoring kidneys prior to transplantation [36].They combined an RGB depth camera for path planning with a swept source OCT for real-time pose adjustment.Both coordinate systems also needed to be registered mutually.Instead of acquiring multiple 3D-FOVs, the robot travels along the sample surface while acquiring 2D B-Scans.In postprocessing, the B-Scans are getting concatenated and adjusted by the robot's position to gain a large-area OCT image of the kidney.Within 20 minutes, a total FOV of 113.9 × 33.8 mm 2 was covered resulting in an imaging speed of 0.032 cm 2 /s.Sprenger et al. introduced a system that relies solely on the data obtained by the OCT without any supplementary spatial information [35].To align the OCT probe with the surface, a low-resolution OCT scan was performed initially at each position to determine the sample's surface position.Their implementation only works for planar samples as it does not consider the sample's orientation for pose adjustment.Stitching of the individual FOVs is completed in postprocessing by using the robot's recorded position.
An example of precise online pose adjustment to a moving subject was shown by M. Drealos et al. and P. Ortiz et al., where they combined machine vision sensors with a robotic arm to align an ophthalmic OCT probe to the eye of freestanding individuals.Although their goal is not to perform large-area imaging, it shows the capabilities of robotic systems combined with OCT probes [39,40].
We present a calibration-free large area robotically assisted OCT (LARA-OCT) system that uses a 7-DOF lightweight collaborative robot and a custom-built swept source OCT (SS-OCT) system with a Fourier domain mode locked (FDML) laser that has a 3.3 MHz sweep repetition rate.Our integration relies solely on OCT as a probe alignment sensor and imaging tool, eliminating the need for additional machine vision systems (e.g., depth or RGB cameras) and elaborate coordinate system calibrations.Furthermore, we have implemented an online control for the probe-to-surface orientation based on the surface information obtained by the MHz-OCT to maintain a perpendicular orientation of the probe relative to the sample surface at a specified distance.The subsequent sections will detail the current implementation, evaluate the performance of the probe-to-surface control, and exhibit large area OCT images acquired by LARA-OCT.Finally, we will examine challenges associated with large-area OCT imaging, their corresponding solutions, and future avenues for enhancing the system's speed and versatility.

Setup description
The proposed system consists of a custom-built MHz-OCT system, a custom-built galvanometric mirror-based OCT probe, a 7-DOF collaborative robot system and a single workstation functioning as the inter-system-link which controls the entire imaging procedure.

Robot setup
The robot used is an LBR iiwa 14 R820 (KUKA AG, Germany) collaborative robot and features seven joints with seven degrees of freedom (Fig. 1).The maximum payload is 14 kg, the maximum reach is between 800-820 mm, and the position repeatability is ±0.15 mm.A collaborative robot like this offers torque sensors in each axis, facilitating force-sensitive motion control.This feature enables the robot to respond to any undesired contact with obstacles, thereby protecting both the subject and the operator from hazards caused by the robot's motion.Furthermore, we implemented several additional redundant safety features, making the system partially up to threefold fail-safe.At the lowest level of safety, continuous monitoring is conducted of the angular speed of the individual robot axes, the cartesian speed of the end effector, and the external forces acting on the robot and the end effector.This is done by the KUKA safety control hardware.Furthermore, the system incorporates two external emergency stops, one for the operator and one for the test subject.A foot switch enables the robot to be placed in hand-guiding mode, thus facilitating the removal of any locked objects.Additionally, an armrest is equipped with an emergency stop mechanism that is triggered in the event of excessive load, such as pressure from the robot on the arm.Furthermore, the armrest cushion can be removed in the event of an entrapment, releasing the arm from a potentially dangerous situation.These efforts are made to operate the system in an open laboratory or open clinical environment to pave the way for clinical trial applications.A KUKA LBR iiwa 14 R820 robot is used to position the OCT probe.Cables to drive the galvanometric mirrors are fed through 3D printed cable guides and strain reliefs.Among others a human arm replica is used as an OCT phantom and placed on a spring-loaded armrest which is used as a limit switch that can trigger an emergency stop.
The robot is connected to a controller (Sunrise Cabinet V10, KUKA AG, Germany) which operates its own development environment (Sunrise.OS 1.17, KUKA AG, Germany) and takes care of all the necessary calculations of forward and inverse kinematics, driving the motor currents or monitoring the safety state.In this environment a custom JAVA application is running which implements a transmission control protocol/internet protocol (TCP/IP) server (Fig. 3).This server receives commands via Ethernet from the workstation with a cycle time of ∼20 ms and converts these instructions into robot specific motion commands.A second TCP/IP-server is running on a different network port as a background task which provides 50 Hz cartesian position updates and further system condition information.This allows us to record robot motion during the imaging procedure and therefore associate a robot pose to each OCT volume.
On the workstation side a custom LabVIEW (National Instruments Corp., USA) based application is responsible for establishing a connection to both TCP/IP servers, commanding the robot, requesting status information, receiving surface height and orientation information from the OCT software, controlling the position of the OCT probe and recording the robot motion.

MHz-OCT system
The light source of our custom-built MHz-OCT is a Fourier domain mode locked (FDML) laser with a fundamental sweep repetition rate of ∼412 kHz and a bandwidth of 100 nm centered at 1310 nm [41,42].To further increase the sweep repetition rate, the sweep is eight-times optically buffered which yields an effective sweep rate of ∼3.3 MHz [43].After buffering, the light is amplified by a boosted optical amplifier (BOA1130S, Thorlabs Inc., USA) so that a power on the sample of >25 mW is achieved.
A custom-built OCT probe is mounted with the help of a 3D printed fixture to the flange of the robot (Fig. 1).A pair of galvanometric scanners (dynAXIS 421, Scanlab GmbH, Germany) is used to steer the light beam across the target.The FOV of a single scan is typically 13 × 13 × 5 mm 3 if using a one-inch spheric lens with a focal length of 50 mm and spectral bandwidth of 100 nm, and 30 × 30 × 20 mm 3 in case of a two-inch 125 mm spheric lens and a spectral bandwidth of 22 nm.The theoretical values for lateral and axial resolution in the case of the one-inch lens are 17.40 µm and 7.46 µm and for the two-inch lens 43.50 µm and 33.90 µm respectively.These datasets are acquired with a balanced 1.6 GHz photodetector (PDB480C-AC, Thorlabs Inc., USA) and an analog-to-digital converter (ATS9373, Alazar Technologies Inc., Canada) which offers an acquisition rate of 4 GB/s and a sample depth of 12-bit.
On the OCT software side, it can be differentiated between three states: (i) live preview of a single B-Scan with ∼180 ms (5.56 Hz) framerate, (ii) image acquisition of the volume and (iii) transferring the data from memory to the hard drive.Because of current software limitations, it is not possible to show a live preview while a scan is performed, or the dataset is transferred from memory to the hard drive.Therefore, it is necessary to implement a structured sequence of instructions for the robot and the OCT system.

Imaging sequence and probe-to-surface control
There are two possible modes to acquire a large area OCT image.First, it is possible to perform a single long scan as an interplay of galvanometer scanner motion in Y-and robot motion in X-direction (long scan mode), resulting in a single elongated image.Depending on the clinical question it might be necessary to perform several individual long scans and stitch them in post to a single large volume.The benefit of this method is the fast acquisition of large areas compared to mosaicking method (see the following paragraph).On the downside, the file size of a single long dataset might exceed already 100 GB making the post processing more computationally intensive.The synchronization of robot motion and galvanometer scanner motion is critically important to achieve a motion artefact-free result.Furthermore, the probe orientation must be precisely and fast adapted to the changing surface structure.Therefore, it is necessary to simultaneously acquire OCT data, transfer the data from memory to the hard drive and update the pose of the robot to align the probe orthogonally to the surface.
Secondly, a large area image can be composed of several individual single-FOV OCT volumes building a large, mosaicked image (mosaicking mode).To raster scan the sample a motion pattern must be defined.Various patterns are conceivable.For example, from the center of a spiral there could be a motion outward that stops as soon as a predefined area has been scanned.Furthermore, a rectangular pattern can be implemented, which is scanned in a meandering manner.In either case the following steps are necessary: (i) align the probe orthogonally to the surface, (ii) perform a single FOV scan, (iii) transfer data from memory to the hard drive, and in the meantime, (iv) move the probe to the next position.These steps are repeated until the desired area of interest is scanned (Fig. 2).After acquiring and storing the data, a two second delay is added to make sure that these steps are safely finished.The time intervals correspond approximately to the measurement from section 4.1.The graphic is not to scale.This sequence is performed by an interaction of OCT software and robot control software, whereby the robot control software not only controls the position of the robot, but also triggers a new image acquisition as soon as the OCT system is ready to scan and store a new volume (Fig. 3).A prerequisite to acquire a volume is that the probe is orthogonally aligned to the surface.
This alignment is done by the probe-to-surface control -a visual servoing based online closed-loop control which orientates the probe orthogonally at a user-defined axial distance to the surface solely based on the information extracted from the OCT preview (Fig. 4(a)).The Since a single B-Scan can only provide one of the orientations, either in X or in Y depending on the scan direction, a second B-Scan perpendicular to the first one is performed.Therefore, we designed a cross-like scan pattern which samples the surface in both directions.From each B-Scan the orientation is determined so that a set of three parameters is gained (distance probe to surface and orientation of the surface along the scan axes).With an update rate of ∼180 ms these sets are then sent from the OCT software to the robot software (Fig. 3).
There, three parallel running proportional gain controllers (P-Controllers), one for each parameter, are orientating the probe orthogonally to the surface (Fig. 4(a)).Since P-controllers suffer from control deviations and oscillations, a tolerance range is defined.Within this range, the probe's orientation is considered orthogonally aligned and the controllers are paused.Please note that the P-Controllers P B (t) and P C (t) share the same control parameters as proportional gain factors, step size limits or setpoints.
Once the alignment is finished, an OCT scan is triggered.While the data is transferred from memory to the hard disk, the probe is moved in a single step to the next position.After storing all the data, the OCT software switches back to the preview mode.Again, a cross-pattern is scanned to acquire the parameter sets.The probe can now be orthogonally aligned again, and a new scan can be triggered.This process is repeated until all individual scans are performed.

Generation of the raster scan motion pattern
Prior to the imaging procedure, the scan positions are determined and subsequently commanded in relation to the current position of the OCT probe.This necessitates the user defining the size of the final FOV in terms of the number of volumes in the x-and y-directions and the step size between each mosaic tile or elongated scan.Consequently, a two-dimensional grid is defined in the world coordinate system of the robot, with the origin located at the base of the robot (Fig. 5).The actual motion of the robot is commanded in the tool center point (TCP) coordinate frame, which is derived from a transformation of the world coordinate frame.This transformation is automatically performed by the Sunrise.OS robot operating system.Additionally, the robot is not commanded to an absolute position in space but relatively to its current position.This enables the probe to be moved along curved surfaces, even though a planar grid is defined, since the grid now depends on the orientation of the probe and therefore on the surface shape.This may result in unevenly spaced volumes.However, if the overlap is sufficiently large (approximately 30%), which is necessary for image registration, the entire surface can be imaged with sufficient overlap.During the whole procedure, rotations around the Z axis of the TCP frame are permitted to keep the x-and y-axes in their corresponding planes.

Image registration and stitching
After acquiring all individual images, the raw data are processed, and the en-face projections are extracted to 16-bit TIFF (tagged image file format) files.Afterwards the files are loaded into Microsoft's image composite editor (ICE 2.0.3,Microsoft Research, USA).A structured panorama is created with 30% overlap and a serpentine image order for a meandering OCT imaging sequence.The same procedure is also applied to stitch individual B-Scans.

Evaluation of probe-to-surface control performance
A step response for each P controller is performed as a first evaluation of the probe-to-surface control performance.A 6-DOF hexapod (H-820, Physik Instrumente GmbH & Co. KG, Germany) is placed under the OCT probe (Fig. 5).The repeatability of the hexapod with ±0.5 µm is smaller than the repeatability of the robot arm (±100 µm) and the theoretical resolution of the OCT in lateral (∼18 µm) and axial direction (∼8 µm).The hexapod is controlled via a custom Matlab (MATLAB 2022a, The MathWorks Inc., USA) application and a serial RS232 link.For each probe-to-surface controller a corresponding step pattern of the hexapod is programmed consisting of 10 consecutive steps.A step size of ±1 mm is used in the positive Z direction of the hexapod (Z hex ) to test the P Z -controller and ±15°in U-and V-directions to test the P B -and P C -controller.The pause between each step is 10 seconds (Fig. 6-8, gray lines).The velocity of the hexapod is set to its physical limit of 20 mm/s.
On the flange of the hexapod a piece of cardboard is fixed to get a flat surface profile when performing OCT imaging.Initially a reference drive for both the robot and the hexapod is performed.The robot is then automatically moved to its starting position, where the Z-axes of the tool center point (TCP) Z TCP and the world coordinate Z W are aligned parallel to each other like the pose in Fig. 1.Then, the probe is brought closer to the hexapod until the surface of

Step response of translation in Z direction
The step response in Z-direction is shown in Fig. 6.The blue curve shows the actual position of the hexapod surface in the OCT image and the gray curve shows the target position recorded by The Z-controller can counteract the fast height change of the hexapod.Within ∼1.5 s after a 2 mm step, the target position can be restored.Minimal oscillations with an amplitude of ∼25 µm follow, which are intercepted by the tolerance range condition.Furthermore, the Z-controller is decoupled from the B-and C-controller.This means that the initial alignment of the sample is orthogonal to the surface.In addition, the robot moves exactly in the direction of motion of the hexapod.By varying the proportional gain parameter K c,Z , the controller exhibits strong oscillations on the one hand (K c,Z ≫ 0.0020) and a slow response time on the other hand (K c,Z ≪ 0.0020).All observations correspond to the typical behavior of a simple proportional gain controller.

Step response of rotation in C direction
In the next step, the C-controller is tested separately.Here, the hexapod performs a rotation in U-direction with an amplitude of ±15°.The proportional gain parameter K c,Z is taken from the previous experiment and the K c,B parameter is varied again until a compromise between high control speed and minimum oscillations is found.Figure 7 shows step responses derived from a combination of K c,Z = 0.0020 and K c,C = 1.40.

Fig. 7.
Step response of the C-Controller after several steps in U direction with a rotational step size of ±15°.
From the green curve, it can be seen that the C-controller can follow the rapid change of the hexapod.During the standstill of the hexapod, the setpoint can be reached for a short moment until the hexapod resumes the motion.Smaller oscillations are visible during the readjustment.In contrast to chapter 3.1, a rotation in C-direction influences the orientation of the surface in B-direction more strongly but still within the tolerance range of ±1°.A readjustment is not necessary.
However, in contrast to the isolated Z-controller characterization, a strong influence of the motion of the hexapod in the displacement direction (here U) on another dimension (here Z) can be detected.This may come from a mismatch of the robot's TCP position and the hexapod's pivot point.Initially, the TCP of the robot may not match the pivot point of the hexapod motion in the Z direction.This leads to smaller angular errors during the alignment, which also affect the height of the surface in the OCT image.The main part of the strong influence comes from the fact that the position of the TCP and the pivot point of the hexapod do not coincide in the X-Y plane.When the hexapod is moved, it causes the FOV on the hexapod surface to shift and thus the center of the fitted straight line through the surface in the OCT image.

Step response of rotation in B direction
Based on the previous measurement, a mismatch of the TCP position and the pivot point of the hexapod results in a stronger coupling between the controllers.Therefore, when investigating the B-controller, an attempt is made to align the two points (TCP and pivot point) on top of each other as closely as possible.The measurements from section 3.2 are also carried out here but adapted for the B-controller.It can be clearly seen that the coupling between the controllers has been significantly minimized.The second orientation is still within the tolerance range and does not need to be readjusted.The performance of the B-controller is comparable to that of the C-controller.This was expected since there were no differences in the control parameters or the controlled system.The alignment of the TCP and the pivot point is still not perfect, but this is not relevant for the further course of the work.

Control performance conclusion
The characterizations performed show that the controllers are separately capable of repeatedly aligning the probe orthogonally to the surface.It was also shown that the position of the TCP and the pivot point of the hexapod have an influence on the coupling between the controllers.This is of particular interest for dynamic samples.For static samples, this should not lead to any problems.Here, care must be taken to ensure that the TCP position is equal to the focus of the OCT and thus on the desired line in the OCT image.In this performance evaluation, a flat surface was selected as the sample to isolate the probe-to-surface control from errors in the surface detection algorithm.It should be noted that the data presented do not necessarily reflect real-world scenarios, given that surfaces, particularly in dermatology, are usually non-planar and a possible trembling of a subject body part is not isolated in one direction.Also, a line fit into the OCT surface might not be sufficient to represent the orientation of the sample.
The OCT preview's current frame rate of 5.56 Hz may be insufficient for rapidly moving objects, necessitating a higher update rate to ensure rapid readjustment of the surface.However, since the objective of the work presented in this paper, is not to track the subject's motion, but rather to align the probe orthogonally to the surface for stationary samples with a slight tremor, the update rate is sufficient.It is possible to optimize the frame rate by enhancing the preview processing pipeline by switching to a resource-efficient programming language.In the future we will interface our 4D live low latency volumetric imaging software with the robot control application to additional also track patient motion [44].
The dynamic interaction of the three controllers can also be seen in Visualization 1.Here, the OCT probe is aligned to a planar surface that is moved arbitrarily at different velocities by hand using a ball joint.

Large area imaging
The imaging modes described in section 2.3 will be tested in more detail in the following.For this purpose, the mosaicking mode with a spiral pattern is tested first, with which the surface of a 3D sphere with a rough surface is raster scanned.Second, the long scan mode is used without probe-to-surface control to demonstrate the imaging speed capabilities of this method.An entire human hand is scanned here (see also Visualization 2).Imaging, scanning and robot motion parameters can be taken from Table 1.All in vivo experiments were conducted on a voluntary basis by experts of our group and approved by the Ethics Committee of the University of Lübeck (including laser and robot safety).

Mosaicking mode
The mosaicking mode can be used in conjunction with the probe-to-surface control and follows the imaging procedure of Fig. 2. To demonstrate the feasibility of a solely MHz-OCT based distance sensor, the rough surface of a 3D printed sphere with a diameter of 20.5 cm is imaged up to a total FOV of ∼52 × 52 × 4 mm 3 (Fig. 9).The robot is initially guided by hand to the apex of the sphere until the surface is within the axial imaging range.Afterwards the imaging process starts, and the robot begins to automatically align the probe orthogonally to the surface.Once the alignment is done, an OCT scan is triggered and after finishing the robot moves to the next position in the TCP coordinate system.Thereby the robot follows a spiral pattern outward for 36 individual scan positions.The recording of the motion path can be seen in Fig. 9(b) as a red line and the gray tiles indicate the position and orientation of the OCT surface (the green tile indicates the first OCT scan).The whole imaging procedure takes about 13 min 30 s with a total scanning time of 43.64 seconds resulting in an imaging speed of 0.03 cm 2 /s.The total data size is about 322 GB.In Fig. 9(a) the stitched large-area image is shown as an en-face projection.Stitching artifacts are not visible but are also hard to see because of the irregular structure of the surface and the blending of adjacent and overlapping image patches.It cannot be completely ruled out that there are stitching errors.The same applies to the stitched B-scan shown in Fig. 9(d).Nevertheless, the robot can move the OCT probe along the curved surface without the need for additional sensors, complex path planning or coordinate system calibrations.Both the height and the angle to the surface can be controlled.The spiral pattern proved to be particularly practical, as the surface could be recorded without prior knowledge of the dimension of the region of interest (ROI).It is also possible to begin imaging from the center of the region of interest, e.g. of a skin inflammation or a material defect, until a sufficiently large area has been scanned.With 0.03 cm 2 /s the imaging speed is rather slow, due to the necessary time to store each of the 8.95 GB large volumes to the hard drive that may be optimized in further development stages.

Long scan mode
As a first demonstration of the imaging speed capabilities of our implementation the long scan mode is demonstrated without probe-to-surface control.With future software implementations it will be possible to simultaneously acquire and store volumes while previewing live OCT data and extracting surface information.To find the correct timing of scan parameters and the velocity of robot motion, several parameter combinations were tested and a cartesian velocity of 70 mm/s was determined as optimum [29].The sweep bandwidth of the laser was reduced to 22 nm to gain axial imaging range to account for the uneven surface of the hand and fingers.The axial imaging range is approximately 18.6 mm.Starting of the home position of the robot where the probe is orthogonally aligned to the workbench, the robot is moved by hand guiding the robot to the back of the hand until the surface is within the axial imaging range.The robot was then placed slightly outside the ROI to have sufficient travel distance available for acceleration.Images taken during the acceleration and deceleration phase of the robot are not scanned equidistantly, which leads to distortions in the image.The measurement is then started, and the robot moves along the hand in a meandering motion.Images are captured both on the forward and return path.In total 6 adjacent and by 33% overlapping elongated single FOVs were captured in a meandering motion pattern and stitched together in post (Fig. 10).The overall acquisition time was 155 s resulting in ∼25.8 s per single volume including an effective scan time of ∼1.6 s and ∼24.2 s to store the volume to the hard drive.The merged FOV is ∼140 × 170 × 20 mm 3 resulting in an imaging speed of ∼1.54 cm 2 /s.In Fig. 10(a) the stitched averaged en-face projection over all 600 depth layers is shown.The stitching result shows few artifacts.A noticeable lateral misalignment is visualized in Fig. 10(b).Due to the reduced sweep bandwidth and the sparse sampling in the direction of robot motion the axial resolution is not sufficient to resolve deeper skin layer boundaries (Fig. 10(d)).Still the en-face projection is showing good detail.A 3D rendering of the index finger's first joint is shown in Fig. 10(c).Our current implementation can currently only merge 2D datasets.Therefore, it is not possible to render overlapping single FOVs in 3D.
In comparison to the mosaicking mode, the long-scan mode is more than 50x faster but also samples the surface more coarsely.

Advantages, challenges and their solutions
Derived from the previous chapters we can see several advantages of LARA-OCT over comparable implementations [34,36].First, the imaging procedure is faster by 2-3 orders of magnitude due to the use of a MHz-swept light source.For dermatological applications, the increased imaging speed facilitates the data acquisition of more comprehensive and precise diagnostic information from larger areas of the skin.This can be further increased by using even faster light sources [45].MHz-OCT imaging also allows in-vivo imaging without significant motion artifacts where patients can lie or stand in a comfortable position without immobilization.This allows longitudinal in-vivo monitoring of dynamic inflammatory skin conditions over time.In non-destructive testing it can save an enormous amount of time needed to detect defects or hidden impurities on a large scale.The operator does not have to focus solely on small ROIs but can instead image over a larger area, detect potential hidden defects or, in the case of skin imaging, inflammations that would otherwise remain undetected.Further it should be possible to also use the imaging modality as an intrinsic machine vision sensor to align the probe to any shaped surface.It does not require any additional machine vision systems or coordinate system registration.
Taking this a step further, with a 50 µm spot spacing, a sweep rate of 3.3 MHz, and an adult human body surface of 1.8 m 2 , it is possible to scan the entire body surface in less than 4 minutes.This may be difficult to achieve for various reasons, but it demonstrates the capabilities of our approach.A more realistic scenario is to image a single side of a forearm of approximately 405 cm 2 with a spot size of 15 µm and a sweep rate of 3.3 MHz in less than one minute (∼7.45 cm 2 /s).We believe that large area OCT imaging is only feasible with MHz-OCT.However, working with MHz-OCT and large-area images containing millions of A-Scans poses several challenges, which will be analyzed in the subsequent discussion.

Data acquisition and management
The data collected in chapters 4.1 and 4.2 yielded over 300 GB of raw data.Expanding the large-area FOV would potentially increase the raw data size to terabytes.The post-processing of raw data takes significant time and necessitates high-end computing to obtain the final image within a reasonable period of time.To mitigate the data size problem, one can reduce the number of A-Scans per volume in exchange for less densely sampled volumes if high-density sampling is unnecessary.However, this is not always desirable, particularly when handling biological samples.Conversely, the prices of hard disk drives per terabyte are decreasing, accounting for a negligible proportion of the total cost of the LARA-OCT engine.

Image registration and stitching
Several registration algorithms were evaluated during our experiments.Feature-based methods in general did not yield satisfactory results.OCT images, particularly B-scans, have limited features that can be correlated between two volumes.Intensity-based methods, which rely on correlation, yielded better results mainly on en-face data.Even in this case, B-scans were challenging to correlate.The processing times and resource requirements were also high due to the large amounts of data.One solution may involve utilizing the position data from the robot, as has already been demonstrated before [36].This data could be employed to roughly align the volumes, and then cross-correlate the individual sub-volumes of overlapping areas with one another.

Data presentation and interpretation
As previously stated, the substantial data quantities lead to longer processing times.Additionally, there are difficulties in registering and stitching the data.For instance, capturing an image of a forearm that comprises 50 individual volumes and 100,000 B-scans requires considerable time to process.We can envision a scenario where the data is initially captured and registered in en-face projection, and the user chooses a designated section from which a composite volume is generated.Artificial intelligence could support this process by automatically analyzing the entire large-area volume and, depending on the application, search for signs of inflammation, cracks, or impurities and marking them accordingly.

System costs
The LARA-OCT is designed for use in clinical settings with patients suffering from skin diseases.One of the primary challenges clinical institutions face in procuring medical equipment is the high cost of acquisition and maintenance.The total cost of the system is mainly composed of the MHz-OCT system (∼60,000 € -100,000 €), the robot (∼50,000 €), and a powerful measurement computer (∼5,000 €).The LARA-OCT is not a standalone system; rather, it is composed of modules.This modular design allows for integrating a possible existing MHz-OCT system with a newly acquired robot.Also enables the two components to be utilized independently of each other.This approach facilitates cost distribution across different cost centers.Additionally, using a less expensive robot, with a price range of 10,000 € to 25,000 €, is a viable option.Given that industrial robots are designed for long, unobstructed service life, the LARA-OCT system should not require significant maintenance costs.

Summary
We have devised a technique that employs MHz-OCT and a seven-axis collaborative robotic arm to overcome the constrained lateral FOV of a single OCT image.Our approach aligns the OCT probe's position perpendicularly on the sample using only the surface information from the OCT signal, without additional machine vision systems.This negates the necessity for initial registration of varying coordinate systems.The system has two acquisition modes, namely, mosaicking mode and long-scan mode.The long-scan mode combines robot motion and galvanometer scanning to achieve an imaging speed of about 1.54 cm 2 /s.Storage pipeline delays currently constrain this speed.The long-scan mode imaged an area measuring 140 × 170 × 20 mm 3 within 2.5 minutes.Theoretically, the system can densely image a forearm in less than a minute or image an entire body with a larger spot size in under four minutes.Here we presented high-quality in vivo OCT imaging of an entire human hand at a 3.3 MHz A-Scan rate.The system allows initial hand positioning of the robot mounted scan head to select the scan region on the volunteer.
The next step is to combine our robot software with our real-time OCT software [46][47][48].This will enable concurrent scanning, probe-to-surface control, and fast storage of volumes on the hard drive.Furthermore, we aspire to advance our system to a clinical setting.
Fig.1.A KUKA LBR iiwa 14 R820 robot is used to position the OCT probe.Cables to drive the galvanometric mirrors are fed through 3D printed cable guides and strain reliefs.Among others a human arm replica is used as an OCT phantom and placed on a spring-loaded armrest which is used as a limit switch that can trigger an emergency stop.

Fig. 2 .
Fig. 2. Timing diagram of the imaging sequence for two consecutive volume acquisitions.After acquiring and storing the data, a two second delay is added to make sure that these steps are safely finished.The time intervals correspond approximately to the measurement from section 4.1.The graphic is not to scale.

Fig. 3 .Fig. 4 .
Fig. 3. Inter-System-Communication Scheme.The LabVIEW-Software is running on a single computer and communicates via TCP/IP with the KUKA sunrise cabinet (Sunrise.OS).Every 180 ms a new surface position is sent to the robot control software which receives in turn every 20 ms new position updates from the robot.

Fig. 5 .
Fig. 5. Setup sketch for the evaluation of the probe-to-surface control performance.(The W index denotes the world coordinate system at the center of the robot's base, TCP the tool-center-point coordinate system at the optical focus and hex the pivot point of the hexapod).

Fig. 6 .
Fig. 6.Step response of the Z-Controller after several steps in Z hex direction with a step size of ±1 mm.

Fig. 8 .
Fig. 8.Step response of the B-Controller after several steps in V direction with a rotational step size of ±15°.

Fig. 9 .
Fig. 9. (a) Large-area OCT image of a 3D printed spherical phantom composed of 36 individual OCT scans resulting in a total FOV of 52 × 52 × 4 mm 3 .(b) Recorded path of the robot motion (red) with OCT-Scans (gray tiles, with starting OCT scan in green).Please note that the tiles do not accurately depict the FOV of individual scans for enhanced visual clarity.(c) Photograph of the 3D-printed sphere surface.(d) Stitched depth scan composed of 6 individual B-Scans.

Fig. 10 .
Fig. 10.(a) Large-area OCT image of a volunteer's hand composed of 6 × 1 long scans with a total FOV of 140 × 170 × 20 mm 3 .The en-face is projected by averaging all 600 depth layers.Y TCP is the motion of the robot and X SCAN denotes the scanning direction of the galvanometric scanner.(b) Excerpt from (a) with visible lateral stitching artifacts indicated by the arrows.(c) 3D visualization of the first joint of the index finger.(d) 16-times averaged B-Scan along the entire hand and middle finger.