In Vivo Feasibility Study: Evaluating Autonomous Data-Driven Robotic Needle Trajectory Correction in MRI-Guided Transperineal Procedures

Abstract

This study addresses the targeting challenges in MRI-guided transperineal needle placement for prostate cancer (PCa) diagnosis and treatment, a procedure where accuracy is crucial for effective outcomes. We introduce a parameter-agnostic trajectory correction approach incorporating a data-driven closed-loop strategy by radial displacement and an FBG-based shape sensing to enable autonomous needle steering. In an animal study designed to emulate clinical complexity and assess MRI compatibility through a PCa mock biopsy procedure, our approach demonstrated a significant improvement in targeting accuracy (p<0.05), with mean target error of only 2.2 ± 1.9 mm on first insertion attempts, without needle reinsertions. To the best of our knowledge, this work represents the first in vivo evaluation of robotic needle steering with FBG-sensor feedback, marking a significant step towards its clinical translation.

I. INTRODUCTION

A. Clinical motivation

Prostate cancer (PCa) stands as one of the most widespread cancers, comprising 14.7% of all new cancer cases in the U.S. The prevalence of PCa highlights its significant impact, affecting approximately 12.9% of all men over their lifetime, as reported by the National Cancer Institute [1]. Transperineal needle placement is a commonly employed technique for PCa diagnosis and treatment. It involves using long 16–18 gauge needles inserted into specific areas of the prostate through the patient’s perineum. Needle placement typically occurs under image guidance, and the clinical outcomes heavily rely on the precision of needle placement. In contrast to other medical imaging modalities, Magnetic Resonance Imaging (MRI) is favored due to its unique capability to visualize both the needle and PCa tissue, enabling more accurate targeting [2].

The conventional MRI-guided transperineal needle placement is performed by a physician who manually inserts the needle through the patient’s perineum, usually guided by a grid template. In MRI-TRUS fusion [3], the needle placement is verified from ultrasound feedback while for in-bore MRI guidance [4], a confirmatory MRI scan is conducted to ensure accurate placement. Frequently, interactions between the needle and tissue layers can cause deviations from the planned trajectory, resulting in targeting errors. An incorrect needle placement is usually corrected with a complete needle retraction and reinsertion from an adjacent point on the grid template. This process can be time-consuming, thereby prolonging the duration of the clinical procedure and subsequently increasing the associated costs. Additionally, repeated needle insertions may increase the risk of complications, such as infection and hematoma, and cause displacement of the prostate from its original position, further complicating accurate targeting.

Consequently, the development of a robotic system capable of simultaneously correcting needle trajectory during its insertion has the potential to reduce the necessity for repeated needle reinsertions during transperineal procedures, thereby enhancing clinical outcomes for PCa patients

B. Related work

Previous research has explored various needle-guidance systems aimed at enhancing the accuracy of MRI-guided transperineal needle placement [4–7]. These devices were designed to supplant the grid template by precisely aligning a needle guide with an MR-identified target from outside the patient’s body, thereby eliminating the drawbacks associated with the gaps between the grid template holes. Similarly to their corresponding MRI-guided traditional procedures with needle grids, these guidance systems are placed in front of the patients perineum, with the base of the device fixed to the MRI table, and have fiducials that allow for image registration. Placement planning is then performed under MRI guidance to determine the desired needle guide position, and the needle is then inserted through the guide into the patients perineum in a straight trajectory. However, achieving consistent in vivo needle placement accuracy remains a challenge largely due to needle deflection caused by various factors, such as asymmetric cutting forces at the needle tip, friction between the needle and tissue, puncture through membranes, and tissue deformation induced by adjacent anatomical structures [8, 9].

Several approaches have been proposed to model and predict needle deflection during soft tissue insertion [10–14], yet their translation to clinical practice remains difficult due to tissue and anatomical variability among patients. Moreover, these models often require intensive computation, limiting their intraoperative use and making them better suited for pre-procedure planning.

Another significant open challenge consists of steering the needle back onto the planned trajectory once it deviates from the target. Several robotic systems were proposed to correct needle deflection during insertion in soft tissue, with the most common approaches involving the use of specially-designed beveled flexible needles [15, 16], or active needles [17–19], though obtaining regulatory approval for such needles presents a major barrier to clinical translation.

Some studies focused on correcting trajectory deflection using base rotation of conventional beveled-tip needles [20, 21] to minimize regulatory requirements. This approach uses the asymmetric reaction forces of the tissue against the bevel to guide the needle in the direction of its tip. Nevertheless, recent studies [22, 23] suggest this approach’s dexterity might be insufficient to compensate for the extent of deflection observed in prostate transperineal biopsy procedures [9].

An innovative alternative involves leveraging radial displacement of the needle shaft (Fig. 1). In this method, a robotic needle guide shifts the needle shaft radially during insertion to steer the needle tip within the tissue. This technique enables correction of deflection for both symmetric and beveled tip needles without the need to fully retract and reinsert the needle. Needle radial displacement was experimentally proven to present increased needle dexterity when compared to needle base rotation [22, 23], but so far, very few works have explored this concept, partly because the proposed methods require detailed modeling of needle-tissue interaction forces and tissue mechanical properties to steer the needle as intended.

Fig. 1.

Schematic with sagittal view of the steering strategy: The needle guide is initially positioned along a straight trajectory towards the desired target. (1) When the needle is pushed forward, it deflects due to needle-tissue interaction forces, causing the tip to deviate from the intended trajectory. (2) Then, the robot displaces the needle guide radially (X and Z directions) to correct the tip deflection, and (3) the needle is pushed forward again to reach the target. This sequence of needle and guide displacements can be used to correct tip deflection as the insertion progresses.

As an alternative, we have recently developed a cooperative data-driven approach for closed-loop needle steering with radial needle displacement [24]. Phantom experiments demonstrated significantly improved targeting accuracy when compared to unassisted insertions. The method combines the advantages of a data-driven approach, which does not rely on a priori modeling of needle-tissue interaction, with increased steerability achieved through the radial displacement of the needle shaft. Moreover, it is applicable to both symmetric and beveled needles and also simplifies the actuation requirements to only translational degrees-of-freedom (DOFs) for needle steering, resulting in a more compact and simpler mechanism for the robotic device.

Despite the promising phantom results, our previous implementation had a few technical shortcomings that prevented the clinical translation of the technology. It required an external electromagnetic (EM) tracking system for accurate needle tip localization feedback. The integration of additional external devices in the workspace, whether it is an EM field generator, a camera, or an optical tracking system, always introduces significant challenges within the clinical environment due to their physical dimensions and spatial requirements, particularly in pelvic interventions. Furthermore, the use of a non-MR-compatible robotic platform and the EM tracking system prevented us from testing the approach in vivo under MR-guidance, which is crucial for obtaining the detailed anatomy of the subject.

C. Contributions

In this study, we integrated a novel fiber-Bragg-grating (FBG)-based shape-sensing needle and integrated to a new 3 DOF MR-conditional robot to achieve in vivo needle trajectory correction validated in the MRI gantry. FBG technology eliminates the requirement for external sensors, such as EM field generators or optical trackers. This significantly reduces the potential for physical interference with imaging devices in clinical settings. Additionally, due to the use of optical fibers in FBG, it exhibits inherent compatibility with most imaging modalities, including MRI. FBG sensors have been used in numerous other studies for sensing the shape of needles [25, 26], catheters [27], and other flexible instruments [28]. They not only provide accurate shape sensing measurements with a small footprint, but are also able to provide fast tracking of the tip pose to be used as closed-loop feedback in needle trajectory control [29, 30].

We also introduce DANTEC (Data-driven Autonomous Needle TrajEctory Correction), an autonomous closed-loop trajectory control for correcting the needle to follow a planned straight-line trajectory. The DANTEC strategy relies on data from the needle tip and robotic guide to adapt a local estimation of the needle model, which is then used it in a model predictive control (MPC) framework to adjust needle trajectory. In the context of needle insertions into real tissue surrounded by multiple anatomical structures, our hypothesis is that our closed-loop strategy will result in better targeting accuracy when compared to regular needle insertions without trajectory correction. To validate DANTEC, we conducted an in vivo study using a live porcine model in a mock single-needle transperineal prostate procedure. Compared to phantom, an animal study presents less controlled conditions, incorporating challenges like diverse and unknown tissue properties and patient-specific anatomy that closely mimic real clinical scenarios. By evaluating DANTEC in an animal model, we aim to not only validate its effectiveness and adaptability under realistic conditions but also to obtain a more precise assessment of its potential for future clinical translation. To the best of our knowledge, this is the first work to perform in vivo closed-loop robotic needle steering using FBG-sensor feedback inside the MR bore.

II. METHODS

A. DANTEC trajectory correction

The needle trajectory correction approach assumes the placement of a robotic needle guide in front of the patient’s perineum, replacing the grid template commonly used for conventional transperineal needle placement. The robotic guide is initially aligned with the target, establishing a straight-line planned path from the guide to the goal. The robot utilizes its horizontal and vertical degrees of freedom (DOFs) to radially move the guide, enabling the steering of the needle to correct its deflection during insertion.

In this work, we expanded upon our data-driven strategy presented in [24] to accommodate fully automated needle insertions. The needle insertion is performed using the robot’s third DOF (Y-direction) to enhance placement accuracy (Fig. 2). While correct depth placement is essential for the efficacy of treatments such as ablations or brachytherapy, in applications like biopsies, the insertion depth error is less critical, as the specimen is obtained from a 1520 mm section along the needle axis. For such procedures, needle insertions could still be performed cooperatively with manual insertion by the operator, provided the system continuously monitors the current insertion length for accurate control feedback.

Fig. 2.

Schematic of the DANTEC trajectory correction approach, with closed-loop feedback of the needle tip provided by FBG-based shape sensing, and using automated insertion steps with a 3 DOF linear stage MRI compatible robot. The inset shows the needle tip coordinates and angles.

In this study, we took a stepwise strategy, where the needle is inserted an additional 5 mm at each step. This stepwise strategy aims to eliminate errors unrelated to the DANTEC approach, such as system latency, thereby allowing for a more precise evaluation of our method. In addition, the displacement of the needle guide for both X and Z directions (Fig. 2) was limited to a safe range of 6 mm around the initial entry point, based on clinical observation [9].

The DANTEC trajectory correction combines a data-driven Jacobian estimator and an MPC controller as depicted in Fig. 2. The estimator uses feedback from the rotary encoders attached to the robot motors to obtain the current needle guide position (Φ_k) and from the sensorized needle to obtain the needle tip pose (Ψ_k). Both of these are registered to the MRI and used to estimate a first-order linear approximation model for the expected tip deflection:

$$\Delta {\Psi }_k\approx J_k\Delta {\Phi }_k,$$ (1)

where ΔΦ_k = Φ_k − Φ_k−1 and ΔΨ_k = Ψ_k − Ψ_k−1 are the measured variations of the needle guide and the needle tip at a given insertion step k, and J_k is a Jacobian matrix.

The Jacobian update calculation is based on the Broydens method [31] and inspired by the algorithm proposed in [32]:

$${\widehat{J}}_{k+1}={\widehat{J}}_k+\alpha \frac{\Delta {\Psi }_k-{\widehat{J}}_k\Delta {\Phi }_k}{\Delta {\Phi }_k^T\Delta {\Phi }_k}\Delta {\Phi }_k^T,$$ (2)

where 0 ≤ α ≤ 1 is a constant parameter for the update rate.

The MPC controller uses the updated model to obtain the needle guide positions that will minimize the tip trajectory error over a prediction horizon H. The tip trajectory error is defined to be $e_k=\left[x_{T_k}-x_{T_0},z_{T_k}-z_{T_0},{\theta }_{h_k},{\theta }_{v_k}\right]$, with $x_{T_0}$ and $z_{T_0}$ being the horizontal and vertical coordinates of the target, and $x_{T_k},z_{T_k},{\theta }_{h_k}$ and ${\theta }_{v_k}$ being the horizontal and vertical positions and angles of the needle tip pose at step k. e_k combines position and angle errors with respect to the desired straight-line trajectory, ensuring that the controller aims not only to place the tip on the planned path, but also to align its orientation to point directly to the target.

For faster MPC optimization, it is not necessary to include the insertion depth $y_{T_k}$ in the error definition because the insertion step values are fixed, and only the length of the final step insertion is adjusted to reach the target depth.

We use (1) to predict ê_k for the next H steps, considering J_k to be constant and equal to the last update with (2). The robot control input u_k is obtained by optimizing the sequence of needle guide displacements ${\left\{\Delta {\Phi }_{k+i}\right\}}_{i=1}^H$ that minimizes the objective function given by

$$c_k=\sum _{i=1}^H\Vert {\widehat{e}}_{k+i}w{\Vert }^2,$$ (3)

where w is a weight balancing values in mm and radians.

The control input u_k is given by the first element of the obtained sequence and is applied to the needle guide robot. After it moves, the insertion step k is finished and step k + 1 begins. The cycle is repeated until the last insertion step is performed and the needle reaches its final insertion depth.

B. Needle shape FBG sensor

The FBG-based sensorized needle (Fig. 3) was build with three FBG-inscribed fiber optic cables placed in a trigonal pattern (120° relative angle), each with four active areas at the sensing locations of 10, 30, 65, and 100 mm from the tip of an 18G nitinol MRI-compatible needle (model KIM18/20, ITP GmbH, Germany). A Micron Optics Hyperion optical sensing interrogator (Luna Innovations, Atlanta, GA, USA) is used to gather the FBG signal responses, providing a wavelength resolution of 1 pm.

Fig. 3.

FBG-based sensorized needle with location of the four active areas. The active area cross-section shows the fiber optic sensors (teal) distributed in a triangular layout.

The FBG sensor data is processed by a shape sensing module that utilizes the sensor-based Lie-group theoretic model [33] to derive curvatures and torsion along the needle axis. The curvature values are obtained through the optimization of a cost function, as outlined in [34]. With needle shape measurements, we can determine the position and orientation of its tip relative to its base, which is attached to the robot, thus providing feedback for our closed-loop system.

C. 3 DOF MRI-compatible robot

The MRI-compatible robot used in this study was inspired by a 2 DOF motorized template previously developed and clinically tested [4, 5]. In this new iteration, we extended the robot to 3 DOF to allow for motorized needle insertion. We eliminated the crossbar design of the original device, which drives the vertical and horizontal bars by two pairs of lead screws, and replaced it with a simple 3-axis linear stage (Fig. 4). While the crossbar design helped keeping the device’s footprint small, it often caused jamming due to desynchronization of each lead screw pair. In contrast, each axis of the new 3-axis linear stage is driven by a single lead screw eliminating the desynchronization issue. The new design also minimizes the use of custom-made metallic components and can be built from commercially available components and 3D-printed plastic housing allowing for easier and low-cost maintenance.

Fig. 4.

3 DOF robot for MRI-guided transperineal needle procedures. The 3-axis frames represent the coordinate systems for the robot, the MR-images, and the Z-frame fiducials used for MRI-robot registration.

The lead screw of each axis of the linear stage is made of titanium, and mechanically coupled to an ultrasonic motor with integrated optical encoders (USR60-E3N, Shinsei Corp., Tokyo, Japan). To minimize the footprints of the linear stage and also reduce the RF and magnetic field interference in the MR imaging volume, all motors are isolated from the stage and placed at the rear end of the device (Fig. 4). They are connected to the lead screws via timing belts (for horizontal (X) and vertical (Z) axes) and a telescopic shaft with U-joints (for the depth (Y) axis). The stage for the depth (Y) axis can be disengaged from the lead screw so that the needle can also be inserted and retracted manually at any time. The motors are connected to three drivers (D6060E, Shinsei Corp., Tokyo, Japan), and low-level positioning is achieved by a motion controller (DMC-4183, Galil Motion Control, Rocklin, CA).

As a safety mechanism, the device can only move within the MRI bore under direct visual supervision. This is accomplished by activating the motor drivers using an MRI-compatible pneumatic footswitch (6210-OB, Herga Technology Ltd, Suffolk, UK) operated by the user positioned in front of the gantry. The needle insertion can be promptly interrupted by releasing the footswitch.

In the current study we focused on single-needle procedures, such as biopsies and focal ablations like laser and microwave. The extension of DANTEC to other applications that involve simultaneous placement of multiple needles, such as cryoablation and brachytherapy, is out of the scope of the current work and would likely require adjustments to the current mechanism to enable disengagement of the already placed needle from the robotic guide before proceeding with the next needle insertion.

D. Software architecture

The described system was integrated with a desktop computer (Ubuntu 22.04) running both 3D Slicer (slicer.org), an open-source software platform for medical image processing and visualization, and the ROS2 Humble framework (ros.org). The FBG shape sensing, DANTEC trajectory control, and robot position control components were implemented as independent packages within ROS2 (Fig. 5).

Fig. 5.

Schematic of the implemented system architecture with its hardware (blue) and software (green) components. The robot’s low-level controller and the FBG interrogator communicate with the desktop computer through Ethernet and USB interfaces, respectively. Meanwhile, 3D Slicer and ROS2 components utilize the OpenIGTLink protocol [35] for communication.

E. MR imaging, registration and targeting

In our workflow, MR imaging is used for robot registration, target planning and final placement confirmation, while closed-loop feedback of the tip position is provided by the FBG sensorized needle. For MRI guidance, we used a 3-Tesla scanner (MAGNETOM Skyra 3T, Siemens Healthineers, Erlangen, Germany). The user interface was implemented in 3D Slicer. Both the animal and the robotic system are positioned on a wooden baseboard specifically designed to fit the patient table of the MRI scanner.

Fiducial frame-based registration is used to register the robot to the MRI scanner coordinate system. We adapted the Z-shaped calibration marker (Z-frame) described in [36] to be horizontally fixed to the baseboard and directly under the animal’s pelvic area. This positioning of the Z-frame ensures that both the pelvic volume and the Z-frame are closer to the scanner isocenter, reducing distortion of fiducial artifacts and enhancing registration accuracy. The position of the Z-frame is firmly established and known relative to the robot, which is also securely attached to the baseboard. The Z-frame fiducials are scanned with a T2-weighted turbo spin echo (T2 TSE) sequence, as specified in Table I. The image is imported to 3D Slicer and used for image-to-robot registration using the 3D Slicer ZFrameRegistration module. Calculated once at the beginning of the experiment, the registration remains valid as long as the MRI patient table is not moved.

Table I.

Intraoperative MRI sequences

	Registration	Planning	Confirmation	Anatomy
Sequence	T2 TSE	T2 TSE	T1 VIBE	T2 SPACE
Orientation	Coronal	Axial	Axial	Axial
Matrix	256×256	256×179	320×195	320×291
FoV [mm]	180×180	160×160	195×240	220×220
Pixel spacing [mm]	0.70	0.625	0.75	0.6875
TR/TE [ms]	3000/108	4800/105	3.8/1.51	1700/105
Flip angle	120°	137°	9°	135°
Slices/spacing	20/2 mm	19/3 mm	60/3 mm	72/2 mm

Using an intraoperative axial T2 TSE image (Table I) of the animal’s pelvic region for guidance, the operator indicates the perineum location and selects the desired targets for needle placement through the 3D Slicer interface. These values and the registration transform are sent to the DANTEC control, where they are used to calculate the needle tip target in robot coordinates through the transformation chain (MRI → Z-frame → Robot), with coordinate systems as represented in Fig. 4. A T1-weighted 3D image is acquired using a 3D volumetric interpolated breath-hold examination (3D VIBE) sequence at the end of each insertion to confirm the needle placement. This sequence is the same used during clinical procedures in our institution to verify final needle placement in transperineal procedures, thanks to the good visibility of the needle in T1 VIBE images (for pixel spacing, see Table I). The tip location is defined by manual inspection and segmentation of the needle shaft within all slices from the scanned 3D volume.

F. Experimental protocol

The study protocol (#2018N000059) has been reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) at Brigham and Womens Hospital. A mock biopsy procedures was conducted using a male swine (Yorkshire, 31 kg). The animal was initially administered a topical anesthetic followed by intubation with a cuffed endotracheal tube, which was secured to the snout using flexible rubber tubing. Oxygen was administered at rates ranging from 1 to 3 liters, and Isoflurane concentrations ranged from 1 to 4% before transportation to the MRI room. Vital signs and depth of anesthesia were continuously monitored using a pulse oximeter, end-tidal CO₂ monitor, and respiration volume/rate throughout the procedure. Once positioned in the dorsal lithotomy position on the baseboard, the robot was securely fixed in front of its perineum (Fig. 6). Subsequently, the MRI table was moved inside the bore with the animal’s pelvis centered at the scanner’s isocenter for imaging throughout the entire experiment.

Fig. 6.

(a) Experimental setup with the animal placed inside the MRI bore. (b) MRI-compatible robot fixed in front of the animal’s perineum.

Five targets were chosen within the pig’s pelvic region, ensuring each was reachable via straight-line transperineal trajectories, similarly to the standard grid template approach (Fig. 7b). For each target, two needle insertions were performed in a stepwise manner, with 5 mm increments: one using an open-loop (OL) insertion method, where the robot aligns the needle guide to the planned entry point and then pushes the needle to the desired depth, and the other using a closed-loop insertion (MPC) method employing the proposed DANTEC trajectory control. In the MPC insertion, the FBG sensorized needle provides feedback on the current tip position, allowing the robot to adjust the needle trajectory by radially displacing the needle guide at each insertion step.

Fig. 7.

(a) Pig anatomy manually segmented post-procedure from T2 SPACE MRI. (b) Target planning: all five selected targets are accessible with straight-line trajectories while avoiding the pelvic bone (beige), urethra (light blue), and rectum (pink).

The confirmation image, acquired at the end of each insertion, is later used for the manual segmentation of the needle. This process determines the final needle trajectory within the tissue for comparison with the planned target. The targeting error, namely the Euclidean distance between planned target and final needle tip position, is then calculated to compare the performance of OL and MPC insertions. Given that the T1 VIBE sequence offers a spatial resolution of (0.75 mm, 0.75 mm, 3 mm), the resolution of the targeting error calculated from the needle segmentation is approximately 1.1 mm in the 2D axial plane and 3.2 mm in 3D. The Shapiro-Wilk test is used to verify the normal distribution of the target errors, and then a paired one-tailed t-test is applied to determine if the error in the OL group is statistically higher than in the MPC group. Finally, at the end of the experiment, a higher resolution T2 image (Table I) is acquired for analysis of the animal anatomy.

III. RESULTS

All closed-loop insertions were successfully carried out to the desired depth. However, two of the open-loop insertions (OL2 and OL3, referring to targets 2 and 3) had to be interrupted due to excessive forces at the needle tip, preventing further insertion.

The detailed pelvic anatomy of the animal was manually segmented to identify its main components, such as organs, glands, bones, and muscles (Fig. 7a). When each needle trajectory was registered against the segmented anatomical structures, it became evident that the narrow pelvic anatomy of the pig model led to all insertions experiencing significant resistance forces due to interactions with bone and muscle layers, causing deviations in their trajectories (Fig. 8).

Fig. 8.

Final needle placement in open-loop (OL) and closed-loop (MPC), obtained from manual segmentation in T1 VIBE confirmation images for targets 1 to 5. All 10 insertions had interactions with the pelvic bone. The circles highlight OL2 and OL3 collisions with bone. The arrow points a large deflection in MPC4 due to entering the interface between muscle layers.

While the needle contact with the pelvis surface in both open- and closed-loop insertions likely induced resistance forces causing trajectory deflections, the closed-loop insertions managed to counteract these effects. They maintained a relatively straight trajectory immediately following the bone contact, in contrast to the open-loop insertions, which exhibited more pronounced deviations from the intended path (Fig. 9). The image shows that MPC insertions, despite the deflections caused by anatomical structures during insertion, consistently maintained closer proximity to the planned trajectory and target compared to their OL counterparts. Moreover, the absence of path correction in OL2 and OL3 led to deviations that resulted in collision with the pelvic bone(Fig. 8), preventing the completion of the insertions.

Fig. 9.

Needle trajectories segmented from T1 VIBE images acquired at the end of each insertion. The yellow dots are the selected target locations, with the planned straight-line paths (blue), the open-loop trajectories (green), the and closed-loop insertions (purple). The insertions interrupted due to collision with bone have been projected (grey) for determination of the final needle tip if such insertions could have been performed until the desired depth. In the bottom corners, axial view of the final needle tip locations for all respective open-loop (OL) and closed-loop (MPC) insertions.

For OL2 and OL3, due to the insertions being prematurely interrupted, we projected the needle along a straight trajectory from its last position to the intended insertion depth (Fig. 9). These projections were used for placement error calculations and correspond to the tip’s potential final position had the insertion proceeded without further deflections.

The final absolute placement errors for each insertion are presented in Fig. 10. We can observe that, in general, MPC insertions exhibited smaller errors when compared to OL insertions with same planned target. The mean target error in OL was 6.4 ± 4.1 mm against 2.2 ± 1.9 mm in MPC. It is important to highlight that these results were achieved on first insertion attempts. Nevertheless, the MPC performance was even superior to the 3.7 mm average error of the best needle placement attempts reported in clinical prostate biopsy cases after multiple needle reinsertions [4].

Fig. 10.

3D absolute targeting error in open-loop (OL) and closed-loop (MPC) insertions for targets 1 to 5.

The Shapiro-Wilk tests provided no evidence that the targeting errors were not normally distributed (p=0.09 and p=0.08, for OL and MPC, respectively). The paired t-test showed a statistically significant difference between the two groups (p=0.030), confirming our hypothesis that our closed-loop strategy results in better targeting accuracy than needle insertions without trajectory correction.

IV. DISCUSSION

All closed-loop insertions presented significantly better targeting accuracy than their open-loop counterparts. The only exception is target 4, where both OL and MPC showed similar performance. Analysis of the MPC4 insertion (Fig. 11) revealed that the needle interacted with bone at a depth of approximately 65 mm, causing significant reaction forces. The controller appropriately responded by adjusting the needle deflection radially, and as a result, the needle’s trajectory continued almost straight toward the desired target, with the guide at its maximum safety limits in both horizontal and vertical directions. However, as the insertion progressed, the needle slid between two muscle layers (arrow in Fig. 8) at approximately 110 mm depth, leading to another significant deflection. At this point, having reached its maximum safety limit, the controller was unable to correct the trajectory due to control output saturation.

Fig. 11.

Final needle trajectory for insertion MPC4. Balloons (a)-(c) display axial slices of T1 VIBE at varying depths, overlaid with the anatomy segmentation. Arrows mark depths of 65 and 110 mm. (a) Needle shaft (purple circle) beyond the bone (beige) interaction zone, still aligned to the target (yellow dot) due to MPC correction; (b) Needle shaft passing through interface between muscles (light/dark reds) and the rectal wall (light brown); (c) Needle trapped between tissue layers causing significant deflection. On the right, 3D views with bones (d) and muscles (e) rendered at lower opacity for visualization of the needle crossing interfaces between tissue layers.

Although the pig model provides a scenario much closer to a real clinical procedure than phantom studies, it also presents certain differences from human anatomy. In humans, the perineum can be easily flattened by pressing the template against the skin, in contrast to the more rigid, round-shaped perineum found in pigs, which also have a tail that prevents the needle guide from being positioned as closely. This results in a larger gap between the guide and the needle’s entry point on the skin, allowing for considerably more needle buckling outside the tissue than seen in human transperineal procedures. Additionally, the pig’s skin is much thicker and more resistant to penetration, which increases the likelihood of the needle deviating from the planned entry point before puncturing the tissue. In all insertions, for both OL and MPC cases, we observed random deviations from the planned trajectory at the first insertion steps. We attribute it to the combination of these factors, that cause unpredictable initial tip angulations after skin puncture. This effect is present in human patients [9], but in a smaller degree.

Furthermore, the anatomy of the pelvic region differs significantly, with the pig presenting a more challenging path to the prostate when compared to humans. In a human patient, the pelvis is more widely open toward the perineum making it easier to access the prostate. In contrast, the pelvis of the pig is stretched along the craniocaudal axis compared to human’s, significantly narrowing the passage for straight insertions from the perineum to the prostate. Critical structures, such as the urethra and the rectum are also running along this passage making it even more challenging to navigate through it. Despite our efforts to avoid the pelvis and the rectum during planning, all 10 insertions encountered significantly greater resistance forces due to interaction with the pelvic bone. This interaction caused OL insertions to largely miss the target, whereas MPC insertions were able to compensate for the deflections. However, the controller reached saturation in many of the closed-loop insertions, leaving little or no room for additional correction thereafter. This poses a problem if additional large deflections occur as the insertion progresses, such as observed in insertion MPC4.

We believe that the conditions encountered in this animal study represent a worst-case scenario, unlikely to be found in human procedures, where less challenging targeting situations are anticipated. However, subjecting the control system to such extreme conditions was invaluable in identifying potential limitations before proceeding to clinical trials. While control saturation emerged as a potential issue, there are feasible solutions to mitigate it. The controller steerability could be improved by reducing the gap between the guide and the perineum, in order to prevent the needle buckling in the outside, because this issue reduces the rigidity of the pivoting point that enables the needle radial steering. One could also consider increasing the imposed 6 mm limit, which is derived from typical lateral displacement seen in manual perineal needle insertions and recognized through clinical finding as not damaging to tissue, though it is not necessarily the ultimate threshold. Further investigation is required to establish the highest limit that ensures safety. In clinical applications that employ a rigid beveled needle, a combination of the proposed radial displacement and needle base rotation methods [20, 21] can be investigated as an alternative to avoid augmenting the safety limit. This approach may enhance needle dexterity at deeper insertion lengths, where steerability by radial displacement is known to decrease [23], while avoiding an increased risk of tissue trauma.

There is room for improvement in the FBG shape sensing feedback. In this study, we used a sensorized needle with four active areas along its length, which was sufficient for high-precision shape estimate in phantom insertions. However, in scenarios with many inflections due to needle steering and reaction forces with anatomical structures, an FBG needle with more active areas would better capture the full needle shape and provide more accurate tip position and orientation estimates. A natural next step would be to replace the current shape sensing with distributed sensing using multicore fibers for improved accuracy, benefiting closed-loop performance. Additional enhancements involve leveraging the entire needle shape provided by the FBG sensors, rather than solely focusing on the tip position. Furthermore, FBG sensors’ ability to detect forces could be used to enable the system to detect proximity to or collision with anatomical structures.

The post-hoc analysis of our results revealed a statistical power of 0.68, which warrant further testing with an increased number of targets. In this study, the number of insertions was limited following the experimental protocol approved by the IACUC, which restricts the total number of insertions to protect the animals from undue distress.

V. CONCLUSION

The results validate the use of our data-driven trajectory correction strategy for improving the targeting accuracy of transperineal procedures. The method was evaluated in vivo in scenarios that accurately reflect, or even surpass, the complexities inherent to needle placement in human pelvic anatomy. Our DANTEC approach demonstrated a significant improvement in accuracy. This confirms that our data-driven approach not only ensures adaptability across varied tissue properties and anatomical configurations encountered in real patient cases, thus providing robustness, but also improve the targeting performance.

To the best of our knowledge, this research represents the first in vivo evaluation of MRI-guided robotic needle steering with FBG-sensor feedback. The findings confirm the potential for clinical translation of the DANTEC approach and contribute to the growing body of evidence supporting the clinical application of autonomous trajectory correction in MRI-guided needle procedures.