Toward autonomous robotic prostate biopsy: a pilot study

Prostate cancer (PCa) is the second most common cancer, after breast cancer, and is one of the leading causes of death in men [11]. Early and reliable detection of PCa has a huge impact on the successful treatment of high-risk patients and on avoiding overtreatment in low-risk patients. The most reliable technique to detect PCa and to estimate its aggressiveness is needle biopsy [14]. Biopsies are carried out most often under US guidance. Traditional transrectal biopsy is replaced by the safer transperineal biopsy, which ensures lower infection risks, but may require sedation and must be performed in the operating room. A way to simplify the transperineal procedure and to reduce patient trauma is to use a small number of entry points, from which to reach all the areas of interest; this procedural improvement, however, is highly dependent on the experience of the doctor.

The PROST robotic system aims at reducing human error in biopsy planning and execution. PROST will add a measure of autonomy to the procedure by using Machine Learning (ML) algorithms during the planning phase (i.e., target selection in the pre-operative MR, fusion of MR with real-time US, entry point planning) and in the execution phase by autonomously positioning a needle guide while tracking the position of the US transrectal probe. Nevertheless, the approval of the target position is done by the user and the insertion of the needle is manual so that the doctor has the complete control of the surgical task.

According to the classification proposed in [15], there are six levels of autonomy for medical robots: The PROST prototype we describe here is positioned at autonomy level 1. The device and the human share decisions and actions; the device autonomously provides both cognitive and manual assistance to the human operator. PROST offers cognitive support during target acquisition and entry point planning, whereas manual assistance consists in facilitating the accurate positioning of the biopsy needle. These functions will allow increasing the procedure’s accuracy independently of the expertise level of the operating physician.

Prostate biopsy is commonly performed by urologists as a freehand procedure. The most advanced techniques use a pointing device to orient the needle and correlate its position to the US image; a US transducer is mounted on a stepper that allows for translation and rotation. We review below the main commercial products and some state-of-the-art prototypes related to prostate biopsy:A further study on systematic robot-assisted biopsy was presented by Han et al. in [2]. The authors employed a robot for transrectal US biopsy previously described in [12].

All the systems briefly described above have some automatic functions, whose operation must be integrated with the human operator’s actions. Biopsy planning is done statically (e.g., in Artemis, MonaLisa) through fusion and target selection; the systems that update the prostate’s position dynamically (e.g., Koelis) do not include dedicated hardware to update the planning, relying only on the experience of the user. By integrating ML techniques for real-time prostate segmentation with robotic hardware, the solution we are presenting here will advance the autonomy of the preceding solutions by offering task replanning in response to perceived changes and so improving the accuracy in target selection.

Our system is designed for precise targeting of the prostate under ultrasound (US) guidance (Fig. 1) in transperineal biopsy procedures. PROST integrates target selection, image fusion and biological motion compensation. We show in synthetic tests that PROST allows physicians with little experience with needle biopsy to reach the same level of accuracy as expert urologists. PROST allows reaching the entire prostate gland through just two punctures. These act as pivot points, resulting in a conical configuration of core positions (Fig. 2). Accuracy in orientation is even more important when the number of insertion points is low; instead of requiring excessive skill of the human, we leverage the robot’s potential for mechanical accuracy. The use of a lower number of insertion points can reduce trauma to the patient without sacrificing the accuracy of a template biopsy; it also leads to a quicker execution. We aim to increase the accessibility of high-quality prostate biopsy thorough compatibility with outpatient settings, quick set-up and fast intervention time independently of the expertise level of the operating physician. The use in outpatients’ clinics will potentially allow time and cost reduction in the procedure in comparison with template biopsy, which is currently carried out in the operating room.

In the next sections, we describe the hardware structure of the robot (“PROST robot” section), then the experimental setup (“Tests on synthetic phantom and Tests on anatomical phantom” sections). “Level of Autonomy and Automatic segmentation of the prostate” sections provide an overview of key elements of the PROST software, whose details are outside the scope of this paper, since these parts are subject to continuous updates. “Results” section presents the results of the tests, while “Conclusions and future work” section summarizes the paper and describes our plans for future developments of the system.

A cohort of 10 urologists was involved in the study. Five were experts, with more than 100 prostate biopsies performed, and 5 inexperienced, with only 10-50 biopsies of experience. The time required for each insertion was less than 1 minute. The first test of each urologist took around 10 minutes, whereas the second test took less than 5 minutes, including the time required for the urologist to get used to the robot interface (Fig. 4). Figure 9 shows the workflow of the test on the synthetic phantom: the user manually selects the 9 targets by moving the US probe and the software registers them with the targets extracted from the CAD model. This registration serves only as reference to visually guide the targeting procedure (Fig. 4 bottom row left), in consideration of the symmetry of the phantom.

The workflow of the anatomical phantom test (Fig. 9) highlights how the robotic prostate biopsy works: the user loads the pre-operative MR data, the robotized US probe acquires the 3D US, targets from MR are registered in the robotic reference system, the doctor checks the registration result and chooses the target points. The robot moves accordingly (Fig. 4).

Table 1 reports the test results. In the synthetic phantom test, the points in the US image are different than the points on the registered structure, due to the registration error between the 3D CAD model and the US image. Nevertheless, in this experiment we chose the points from the US image as targets for needle insertion. The reason is that the synthetic phantom tests serve only to measure the accuracy of the robot mechanism and the calibration of the system. When the needle trajectory is distant from the axis of the conic workflow, the targeting error is larger. The average error we obtained for the synthetic phantom, considering all the trials and all the participants, was \( 0.91 \pm 0.72 mm\) in the reference frame of the robot and \( 1.30 \pm 0.44 mm\) in the US image. An obvious fact we noticed was that the error does not depend on the experience of the user, but on the precise preparation of the setup (calibration and identification of the target points).

The results on the anatomical phantom (Table 1) show little difference with the synthetic phantom tests. The main difference in the procedure is that here we chose the targets in the MR image. Even if the targets are visible in the US images, we wanted to replicate the setup of a real targeted fusion biopsy. The MR image is registered with the robot by using the segmentation of the whole prostate gland (“Tests on anatomical phantom” section) and the rigid registration algorithm. In this case, the registration error also accumulates with the previous sources of error, especially with the image calibration error. The average error, computed over all the participants and all trials, was \( 1.26 \pm 0.48 mm\) in the robot reference frame and \( 1.55 \pm 0.34 mm\) in the US image. By using the US image only we have an estimation of the error independent of the robotic system. The higher accuracy of the measurement in the reference frame of the robot is due to the rigid assumption, when if fact the needle may curve due to the bevel effect [5]. Again, we did not notice any significant differences between experienced users and non-experienced users. We also noticed a better accuracy and precision in the synthetic test as compared with the anatomical test (Table 1, last row). This is due to the fact that the in the synthetic phantom the structure was rigid (3D printed) and the targets were smaller. In this case, the needle could stop at the target, while in the anatomical phantom the target structures were soft and the needle could overshoot.

Our results prove the usability of PROST, regardless of user experience and with an accuracy of around 1 mm. Considering that clinically significant lesions tend to exceed 5 mm of diameter, the system allows collecting targeted biopsies for every lesion of the prostate.

In this paper, we presented the first prototype of PROST, a prostate biopsy robot that includes a level of autonomy in target identification, image fusion and needle guidance. PROST can guarantee that high-accuracy prostate biopsies are accessible, thanks to a low-cost solution that works independently of the clinician’s experience.

The PROST robotic system has several potential clinical advantages with respect to current products and laboratory prototypes: accurate targeting—comparable with MR guided biopsy but with the advantage of real-time US guidance and 3D visualization—repeatability of the biopsy for active surveillance by mapping all the locations of biopsy cores back into the MR image on file through robotic registration, standardization of the biopsy procedure regardless of the user’s experience level, less trauma for the patient, and the possibility of combining the robot with other diagnostic and therapeutic devices. The next step will be testing on cadavers, to verify the portability of this system to a clinical setting.

The design of a second prototype will take into account both objective parameters (robot workspace, movements of the mechanism, compatibility with clinical standards, software and hardware requirements) and subjective evaluations (e.g., ergonomics, usability, improved user interface) to reach the goal of a safer and more accurate prostate biopsy.

Particular attention will be paid to the sterilization constraints required of a surgical device. Small parts that are in contact with the needle, such as the robotic arms, will be removable and disposable, while the mechanical part will be detachable from the electronic part for autoclave sterilization. The design of the US probe holder will allow for easy removal and application of a sterile condom onto the probe. The new prototype will also support semi-automatic saturation biopsy through the registration of a prostate atlas and a biopsy scheme with real-time segmented prostate.

We are training the segmentation module on patient data. Real-time segmentation will cope with the movement of the prostate during the biopsy procedure, through 2D contour registration and subsequent tracking of the prostate. The image analysis software will also include techniques to handle prostate deformation and will pave the way for the next level of autonomy.