Training system for neural networks to guide robotic arms in manipulating conduits

Abstract

The invention disclosed herein is entitled "Training System for Neural Networks to Guide Robotic Arm Manipulation of Catheters". This invention provides methods and systems for training machine learning (e.g., neural networks) or other artificial intelligence (AI) models to control a robotic arm to manipulate catheters in a robotic manner to perform invasive clinical catheter-based procedures.

Description

[0001] Cross-references to related applications

[0002] This patent application claims the benefit of U.S. Provisional Patent Application 63 / 271,270, filed in October 2021, the disclosure of which is incorporated herein by reference. Technical Field

[0003] The present invention relates generally to robotic surgery, and more specifically to catheters coupled to a robotic arm that are automatically guided and manipulated by a neural network (NN). Background Technology

[0004] Controlling medical robot systems has been previously proposed in patent literature. For example, U.S. Patent Application Publication 2014 / 0046128 describes a control method applicable to a surgical robot system, which includes: a slave robot having a robotic arm to which primary and auxiliary surgical tools are attached; and a master robot having a master manipulator for manipulating the robotic arm. The control method includes acquiring motion data about the master manipulator, predicting basic movements to be performed by the operator based on the acquired motion data and learning results of multiple movements constituting the surgical task, and adjusting the auxiliary surgical tools according to the predicted basic movements to correspond to the operator's basic movements. This control method allows the operator to perform surgery more comfortably and moves or fixes all necessary surgical tools to or in the optimal surgical position.

[0005] For example, U.S. Patent Application Publication 2021 / 0121251 describes a device for robotic surgery that includes a processor configured to receive patient data from treated patients, receive surgical robot data from each of the plurality of treated patients, and output a treatment plan for the patient to be treated in response to the patient data and surgical robot data. This approach has the advantage of adapting to individual differences between patient and surgical system parameters to provide improved treatment outcomes.

[0006] U.S. Patent Application Publication 2020 / 0297444 describes certain aspects of a system and technique relating to the positioning and / or navigation of a medical device within a lumen network. The medical system may include: an elongated body configured for insertion into a lumen network; and an imaging device positioned on a distal portion of the elongated body. The system may include a memory and a processor configured to receive image data from the imaging device, including images captured when the elongated body is located within the lumen network. The images may depict one or more branches of the lumen network. The processor may be configured to access a machine learning model of one or more lumen networks and, based on the machine learning model and information about the one or more branches, determine the positioning of the distal portion of the elongated body within the lumen network.

[0007] This disclosure will be more fully understood through the following detailed description of embodiments thereof, taken in conjunction with the accompanying drawings, wherein: Attached Figure Description

[0008] Figure 1 This is a schematic diagram of a training system for a neural network (NN) to guide a robotic arm in manipulating a cardiac catheter, according to an embodiment of the present invention; and

[0009] Figure 2 The present invention provides a schematic block diagram and flowchart illustrating a combination of methods and algorithms for training and using a neural network to guide a robotic arm to operate a conduit, according to an embodiment of the invention. Detailed Implementation

[0010] Overview

[0011] Probes, such as catheters, including cardiac catheters, are typically operated manually by a physician using a handheld device (e.g., a handle equipped with controls such as knobs and actuators) to properly position the probe within an organ, such as the heart. For example, to diagnose and treat arrhythmias, a physician may need to insert a multi-electrode catheter into a patient's heart to measure intracardiac signals via the catheter and apply ablation based on the findings. This requires physicians to operate the catheters with high skill, often under demanding working conditions that limit their availability.

[0012] To improve the usability of general catheter insertion procedures, and particularly the usability of cardiac catheterization-based procedures, some embodiments of the disclosed technology envision catheter manipulation performed by a robotic arm operated by a neural network (NN). This NN translates high-level requests from physicians for specific catheter manipulations (e.g., by specifying a target anatomical structure to be approached) into actions performed by the robotic arm operating a handpiece. This handless robotic control requires a set of training data for the NN to train the robotic system, which is currently unavailable.

[0013] The embodiments of the present invention described below provide methods and systems for training machine learning (e.g., neural networks) or other artificial intelligence (AI) models to control a robotic arm to manipulate catheters in a robotic manner to perform invasive clinical catheter-based procedures.

[0014] Some embodiments of the present invention provide one or more training devices, each including a set of sensors attached to a handpiece of a probe (e.g., a catheter) used by a physician to perform invasive procedures. The sensors record the movement of the handpiece, i.e., as its operation causes the distal end of the catheter to translate, orient, rotate, and lock into place. The sensors also track the physician's activation and manipulation of elements of the handpiece (e.g., a knob for deflecting the distal end of the catheter). The term "handpiece" generally refers hereinafter to the proximal portion of the probe handle where controls are located, but "handpiece" may also include a proximal section of the shaft and / or the probe's sheath located near the handle, which the physician may also manipulate. As the physician manipulates the handpiece, sensor readings are recorded, and the corresponding positioning of the catheter in the catheter sensor is also recorded.

[0015] The training equipment of the disclosed training system is used by one or more (usually many) physicians, and the training data obtained from each physician's actions is recorded and assembled into a set of data for training (e.g., teaching) the NN.

[0016] Sensors can be of many and different types, such as magnetic sensors, electrical sensors, and “coded” sensors—that is, those that use encoders. Figure 1 The document describes a detailed implementation of this training system with its specific sensors.

[0017] The training objective is a clinical action, such as reaching a target location, and the NN optimization process (e.g., defined by a loss function) can be achieved by using the minimum number of catheter movements performed via the handpiece or by using the minimum amount of spatial movement at the distal end in each operative step. For example, suppose the catheterization task is to anatomically map an 8×8×8 cm atrial volume with a resolution of 2 mm, i.e., comprising 64,000 voxels. The operator wants the cardiac catheter to move and provides two voxel counts to the network: a starting position (e.g., voxel 15) and an ending position (e.g., voxel 2050). In this implementation, the NN must be trained to find the most recent movement or a series of movements from a given starting location to a given ending location (e.g., it could be two movements from voxel 15 to voxel 350 from the location (e.g., position, orientation, and roll angle), and then a single movement from voxel 350 to voxel 2050, or from voxel 15 to voxel 2050). Once the NN selects the nearest (i.e., smallest) movement, it outputs commands to the robot corresponding to the movement of the handpiece, that is, using control of the handpiece to translate (e.g., advance / retract), deflect, and rotate the distal end of the duct shaft.

[0018] To ensure that the aforementioned movements commanded by the neural network (NN) are sufficiently accurate and consistent in adapting to different patient anatomy, it is envisioned that at least several dozen NN training datasets have been collected. However, since hundreds of such procedures may be executed in a year, the training system can easily acquire very broad and robust training data sets.

[0019] System Description

[0020] Figure 1 This is a schematic diagram of a training system 20 for a neural network (NN) to guide a robotic arm to operate a cardiac catheter according to an embodiment of the present invention. Figure 1 Depicting a physician using a multi-arm Catheter 21 performs mapping and / or ablation of the heart 26 of patient 28 on operating table 29. As shown in illustration 25, catheter 21 includes a multi-arm distal end assembly 40 at its distal end, coupled to electrode 42. During catheter insertion, electrode 42 acquires electrical signals from the tissue of heart 26 and / or injects electrical signals into the tissue of heart. The distal end assembly 40 is further coupled to a magnetic positioning sensor 50, which is configured to output signals indicating the position, orientation, and roll angle of the distal end assembly 40 within heart 26.

[0021] When using sensor 50, processor 34 in console 24 can thus determine the position, orientation, and roll angle of the distal end-effector assembly. For this purpose, console 24 further includes drive circuitry 39 that drives a magnetic field generator 36 placed at a known location outside the patient 28 (e.g., below the patient's torso). Physician 30 can then observe the positioning of the distal end-effector assembly 40 in an image 33 of the heart 26 on user display 32.

[0022] Magnetic positioning sensing methods, such as those used in CARTO manufactured by Biosense Webster, are employed. TM The system is implemented and described in detail in U.S. Patents 5,391,199, 6,690,963, 6,484,118, 6,239,724, 6,618,612, and 6,332,089, in PCT Patent Publication WO 96 / 05768, and in U.S. Patent Application Publications 2002 / 0065455A1, 2003 / 0120150A1, and 2004 / 0068178A1, all of which are incorporated herein by reference.

[0023] In addition to magnetic positioning tracking, during this procedure, an electropositional tracking system can be used to track the corresponding position of electrode 42 by associating the positioning signal acquired by each electrode with the cardiac positioning from which the signal was acquired. For example, an active current position (ACL) system manufactured by Biosense-Webster (Irvine, California) as described in U.S. Patent 8,456,182, the disclosure of which is incorporated herein by reference, can be used. In the ACL system, the processor estimates the corresponding position of electrode 42 based on impedance measurements between each of the catheter electrodes 42 and multiple surface electrodes (not shown) coupled to the skin of the patient 28.

[0024] Physician 30 navigates the distal end assembly 40 mounted on the shaft 22 to the target location in the heart 26 by holding the catheter handle 31 with the right hand and manipulating the shaft 22 by pushing the shaft with the left hand or by using the thumb control 43 near the proximal end of the handle 31 and / or deflecting it relative to the sheath 23.

[0025] As shown in Figure 45, the handle 31 further includes a knob 27 with a locking button 37, which the physician 30 also uses to manipulate the shaft 22. A set of sensors 52-60, described below, are coupled to the catheter 21 at various positions to track the physician's manipulations. In particular, multiple sensors 52-60 record the movement of the handle 31, as well as the proximal segments of the sheath 23 and shaft 22, i.e., when the physician 30 translates and rotates the catheter 21 during catheter insertion procedures. The physician 30's manipulations of the catheter 21 are collectively referred to below as "manipulation of the handpiece," and the purpose of the system 20 is to generate positioning data that enables the training of a neural network (NN) to be used by a processor to guide the catheter using a robotic arm, such as... Figure 2 As stated above.

[0026] For example, components of the handheld device, such as a knob 27 for deflecting (44) (i.e., by variable bending redirection) the distal end of the shaft 22 of the catheter and a button 37 for locking it in place, and a thumb control 43 for advancing or retracting (62) the shaft 22 and rotating (64) the shaft 22, are activated by the physician. As the physician manipulates the handheld device, readings from sensors 52-60 are recorded, as are corresponding readings from sensors 50 on the catheter and / or the advanced positioning current (ACL).

[0027] In the illustrated embodiment, sensor 52 is a magnetic sensor configured to indicate the position, orientation, and roll angle of handle 31. Sensor 54 is an coded sensor configured to indicate the amount and direction of rotation of knob 27. Sensor 56 is configured to indicate the timing and duration of use of button 37. Sensor 58 is a magnetic sensor configured to indicate the position, orientation, and roll angle of the proximal segment of sheath 23. Sensor 60 is an coded sensor configured to indicate the amount and direction of distal or proximal movement 62 of shaft 22. Sensor 59 is an coded sensor configured to indicate the amount and direction of rotation 64 of shaft 22 about its longitudinal axis. The entire recorded catheter position information is stored in memory 38 in console 24 for use during NN training, such as... Figure 2 As stated above.

[0028] Figure 1 The exemplary illustrations shown are chosen solely for conceptual clarity. Additional or alternative sensors, such as sensors 52-60, can be used for positioning. Other types of sensors and conduits can be equivalently employed during training. The sensors can sense one or more movements of the actuator located at the handpiece.

[0029] Contact with the tissue sensor may be mounted at the distal end assembly 40. In an embodiment, the processor 34 is further configured to indicate the quality of physical contact between each of the electrodes 42 and the inner surface of the cardiac tissue during measurement.

[0030] Processor 34 typically includes a general-purpose computer having software programmed to perform the functions described herein. Specifically, processor 34 runs the functions disclosed herein. Figure 2 The software incorporates a dedicated algorithm that enables processor 34 to perform the steps disclosed in this invention, as further described below. The software can be downloaded electronically to a computer via a network, or alternatively or additionally set and / or stored on a non-transitory tangible medium (such as magnetic storage, optical storage, or electronic storage).

[0031] Training a neural network to guide a robotic arm to manipulate ducts

[0032] Figure 2 This is a schematic block diagram and flowchart illustrating a method and algorithm for training and using an NN250 to guide a robotic arm to operate a conduit, according to an embodiment of the present invention. The block diagram and flowchart are divided into two phases: a training phase 201 and an operation phase 202.

[0033] Training phase 201 involves multiple training systems (such as...) Figure 1 The training data consists of two corresponding datasets: (System 20) and / or a collection of training data from multiple procedures executed on the system (such as training system 20).

[0034] • First dataset 236: Probe handheld sensor 233 (e.g., Figure 1 The handheld device with sensors 52-60 in the conduit 21) and

[0035] • Second dataset 242: distal probe readings and catheter positioning readings 240 from within the organ (e.g., the corresponding position, orientation, and roll angle of the distal end assembly 40 of catheter 21 in the heart 26).

[0036] Multiple datasets 236 and 242 are stored in memory (such as memory 38 of system 20) for training NN 250 to become trained NN 255.

[0037] Operational phase 202 involves the use of a trained neural network (NN) in a system including probes coupled to a robotic arm. Such a system is described, for example, in U.S. Patent 8,046,049, which describes a device for use with a manipulable catheter, the device including a thumb control adapted to control, for example, movement of an axis or deflection of a distal tip of the catheter. The device includes a robot with end effectors adapted to be coupled to the thumb control, such as thumb control 43 or knob 27, to advance / retract and / or deflect a distal tip, such as distal assembly 40. U.S. Patent 8,046,049 is assigned to the assignee of this current patent application, and its disclosure is incorporated herein by reference.

[0038] like Figure 2 As shown, during operation phase 202, the physician delivers the desired catheter movement 260. For example, the physician (such as physician 30) may command the mapping of heart chambers and / or ablation within them (e.g., ablation of the pulmonary vein ostium). An neural network 250 trained (i.e., taught 248) using a collection 238 of datasets from numerous heart chamber mapping and / or ablation sessions translates the high-level instructions 260 into actual commands 270 for the robotic arm to manipulate the catheter according to the learned commands.

[0039] Figure 2 The example block diagrams and flowcharts shown are chosen solely for conceptual clarity. This embodiment may also include additional steps to the algorithm, such as a confirmation request step. These and other possible steps have been intentionally omitted from this disclosure to provide a more simplified diagram.

[0040] Although the implementation schemes described herein primarily address cardiac catheter-based applications, the methods and systems described herein can also be used to train any medical robotic arm to operate probes coupled to the robotic arm.

[0041] Implementation Examples

[0042] Example 1

[0043] The embodiments of the present invention described below provide a training system (20) including one or more training devices and a processor (34). Each of the one or more training devices includes a probe (21) for insertion into an organ, the probe including: (i) a plurality of sensors (52-60) located at a handheld part (31) of the probe, the sensors being configured to sense probe manipulation by a physician using the probe to perform an invasive procedure; and (ii) a distal sensor (50) located at a distal end of the probe (21) and configured to indicate the position of the distal end corresponding to the probe manipulation within the organ. The processor (34) is configured to (a) receive probe manipulation data acquired from the one or more training devices using the plurality of sensors (52-60) and the corresponding distal sensor (50) of each training device, and (b) train a machine learning (ML) model (250) using the probe manipulation data to guide a robotic arm to manipulate the probe to perform the invasive procedure.

[0044] Example 2

[0045] According to the training system of Embodiment 1, the sensors (52-60) at the handheld device (31) are configured to sense at least one probe operation type selected from a group of types consisting of: position adjustment, orientation adjustment and roll angle adjustment of the distal end of the probe (21).

[0046] Example 3

[0047] According to any one of Embodiments 1 and 2, the training system wherein the sensors (52-60) at the handheld device (31) are configured to sense at least one probe operation type selected from a group of types consisting of: advance, retraction, deflection, and rotation of the distal end of the axis of the probe (21).

[0048] Example 4

[0049] According to any one of embodiments 1 to 3, the sensor (52-60) at the handheld device (31) is configured to sense the probe operation by sensing one or more actions of the actuator located at the handheld device.

[0050] Example 5

[0051] According to the training system of Embodiment 4, the actuator includes at least one of the distal section of the probe sheath, a thumb control, a knob, and a locking button.

[0052] Example 6

[0053] According to any one of embodiments 1 to 5, the sensor at the handheld device and the remote sensor are one of a magnetic sensor, an electrical sensor, and a coded sensor.

[0054] Example 7

[0055] According to any one of embodiments 1 to 6, the training system wherein the processor (34) is configured to train the ML model (250) to find a minimum number of movements from a given starting location in the organ to a given ending location at the distal end of the probe in the organ.

[0056] Example 8

[0057] The training system according to any one of embodiments 1 to 7, wherein the ML model (250) is a neural network (NN).

[0058] Example 9

[0059] A training method includes inserting a probe (21) into an organ in each of one or more training devices, the probe including a plurality of sensors (52-60) located at a handheld part (31) of the probe, the sensors being configured to sense probe manipulation by a physician performing an invasive procedure using the probe. The probe (21) further includes a distal sensor (50) located at a distal end of the probe and configured to indicate the position of the distal end within the organ corresponding to manipulation of the probe (21). Probe manipulation data is received from the one or more training devices, the data being acquired using the plurality of sensors (52-60) and the corresponding distal sensor (50) of each training device. A machine learning (ML) model (250) is trained using the probe manipulation data to guide a robotic arm to manipulate the probe (21) to perform the invasive procedure.

[0060] Therefore, it should be understood that the above embodiments are cited by way of example, and the invention is not limited to the specific contents shown and described above. Rather, the scope of the invention includes combinations and sub-combinations of the various features described above, as well as variations and modifications thereof, which will be apparent to those skilled in the art upon reading the above description, and which are not disclosed in the prior art.

Claims

1. A training system for a neural network (NN) to guide a robotic arm in manipulating a cardiac catheter, comprising: One or more training devices, each training device comprising: A catheter for insertion into the heart, the catheter comprising: Multiple sensors, located at the handheld portion of the catheter, are configured to sense manual manipulation by a physician performing invasive procedures using the catheter; and A distal sensor, located at the distal end of the catheter and configured to indicate the position of the distal end within the heart corresponding to the manual operation; and Processor, the processor being configured to: Receive manual operation data acquired from the plurality of sensors and the corresponding distal sensors of each training device from the one or more training devices; and In the absence of a physician performing manual intervention, a machine learning (ML) model is trained using the manual intervention data to guide a robotic arm to manipulate the handpiece of the catheter to perform the invasive procedure.

2. The training system of claim 1, wherein, The sensor at the handheld device is configured to sense at least one manual operation type selected from a group of types consisting of: position adjustment, orientation adjustment, and roll angle adjustment of the distal end of the catheter.

3. The training system of claim 1, wherein, The sensor at the handheld device is configured to sense at least one type of manual operation selected from a group of types consisting of: advance, retraction, deflection, and rotation of the distal end of the shaft of the conduit.

4. The training system of claim 1, wherein, The sensor at the handheld device is configured to sense the manual operation by sensing one or more movements of the actuator located at the handheld device.

5. The training system of claim 4, wherein, The actuator includes at least one of the distal section of the sheath of the catheter, a thumb control, a knob, and a locking button.

6. The training system according to claim 1, wherein, The sensor at the handheld device and the remote sensor are one of a magnetic sensor, an electrical sensor, and a coded sensor.

7. The training system according to claim 1, wherein, The processor is configured to train the machine learning (ML) model to find the minimum number of movements from a given starting point in the heart to a given ending point at the distal end of the catheter in the heart.

8. The training system according to claim 1, wherein, The machine learning (ML) model is a neural network (NN).

9. A training method for a neural network (NN) to guide a robotic arm in manipulating a cardiac catheter, comprising: In each of one or more training devices, a catheter is inserted into the heart, said catheter comprising: Multiple sensors, located at the handheld portion of the catheter, are configured to sense manual manipulation by a physician performing invasive procedures using the catheter; and A distal sensor, located at the distal end of the catheter and configured to indicate the position of the distal end within the heart corresponding to the manual operation; Receive manual operation data acquired from the plurality of sensors and the corresponding distal sensors of each training device from the one or more training devices; and In the absence of a physician performing manual intervention, a machine learning (ML) model is trained using the manual intervention data to guide a robotic arm to manipulate the handpiece of the catheter to perform the invasive procedure.

10. The training method according to claim 9, wherein, The sensor at the handheld device is configured to sense at least one manual operation type selected from a group of types consisting of: position adjustment, orientation adjustment, and roll angle adjustment of the distal end of the catheter.

11. The training method according to claim 9, wherein, The sensor at the handheld device is configured to sense at least one type of manual operation selected from a group of types consisting of: advance, retraction, deflection, and rotation of the distal end of the shaft of the conduit.

12. The training method according to claim 9, wherein, The sensor at the handheld device is configured to sense the manual operation by sensing one or more movements of the actuator located at the handheld device.

13. The training method according to claim 12, wherein, The actuator includes at least one of the distal section of the manual sheath, a thumb control, a knob, and a locking button.

14. The training method according to claim 9, wherein, The sensor at the handheld device and the remote sensor are one of a magnetic sensor, an electrical sensor, and a coded sensor.

15. The training method according to claim 9, wherein, Training the machine learning (ML) model involves training the machine learning (ML) model to find the minimum number of movements from a given starting point in the heart to a given ending point at the distal end of the catheter in the heart.

16. The training method according to claim 9, wherein, The machine learning (ML) model is a neural network (NN).

Images