AI-driven autonomous robot thrombectomy in acute stroke 

Project reference:  SIE_10_22
First supervisor:  Thomas Booth
Second supervisor:  Alejandro Granados Martinez

Start date: October 2022

Project summary:  In 2016 level 1a evidence showed that mechanical re-perfusion therapy (thrombectomy) is a highly effective treatment for acute ischaemic stroke within 6 hours and better outcomes are seen the sooner re-perfusion is achieved (1 ,2). The ability to perform a thrombectomy or not within the time window is disproportionately dependent on geographical location (2), given the hours taken for diagnosis at the local hospital and transfer to specialist neuroscience centres.

 

Aim: develop AI-driven autonomous robot thrombectomy to allow local treatment with minimal remote supervision from a specialist neuroscience centre. This would reduce the time gap from diagnosis to thrombectomy. If proven to be safe and effective in clinical trials (beyond PhD), such systems could be located in several geographically strategic sites and would almost certainly have huge impact in clinical outcomes (2).This project will use machine learning algorithms to leverage important components of an autonomous surgical robot, namely object recognition and navigation.

 

Objectives:

  • Learn a compact and multimodal representation of inputs

  • Map gestures and position of tools and investigate multi-stage navigation approaches

  • Design sensorimotor control through physical interactions and update policies in real-time for autonomous navigation

Project description:

Stroke is the second leading cause of death across the world, annually killing approximately 6 million people and the third leading cause of disability (3). Given time and geographical constraints,thrombectomy presents a significant challenge for health services and a fraction of eligible patients are treated by thrombectomy (3,4,5,6,7).

Robotically-assisted neurointervention is a nascent field motivated by improving the procedural “micromovements”through, for example, joystickcontrol (8).Our team are developing a neurointerventional tele-operated manipulator(copying the operator’s movements directly via a remote connection) motivated by the need to overcome geographical challenges preventing timely thrombectomy. There are multiple challenges we are overcoming by building a fit-for-purpose system(8) .In particular, the large amount of delicate manipulation of the tiny catheter (2.1 Fr) and wire (0.014 inch) that occurs during the procedure demand that a major focus of our work relates to haptic feedback which is a major concern in the current joystick-controlled systems (8,9,10,11). To ensure the haptic feedback reliably reproduces contact force on vasculature (i.e. safety), we are now performing in silico experiments using SOFA software to simulate catheter manouevres through the vasculature during thrombectomy. We have built neurovasculature phantoms, responders and controller prototypes and are performing in vitro simulations performed by neurointerventional experts to assess the haptic feedback.

 

The work to date provides the ideal in-house basis for developing AI-driven autonomous thrombectomy. The highly-challenging autonomous execution of tasks and applying cognition for decision making will be enabled by using:

 

  • Multi-modal inputs to learn a compact and underlying representation of the instruments reaching a target through the vasculature(12).

  • In silico simulations to assess how a learning agent can improve the efficiency and generalisation of learning a new task (13).

  • A navigation process within a multi-stage training pipeline from a combination of Hamilton-Jacobi-Bellman policies and reinforcement learning, compared to wall-following approaches (14). Goals will be defined at increments of the vasculature that lead to the targetpoint (15.)

  • Robust sensorimotor control under uncertainty through physical interactions of catheters and guidewires while navigating through the vasculature (16)

  • Hand gestures (17) linked to the position of tools (18) observed through fluoroscopy using spatial and temporal embeddings in combination with transformers or recurrent neural networks.

 

Months 1-3: Review of literature on AI-driven robotic systems; exposure to thrombectomy procedures (King’s College Hospital) and understand tele-operated thrombectomy hardware.

 

Months 4-9: Use the in silico simulation and run thousands of additional simulations of navigating catheter/wire through vasculature during thrombectomy. Clean data and input for an AI model to determine optimal pathway for navigation milliseconds before the anatomy is reached(for example, catheters often cannot pass through the tortuous carotid siphon; we will determine how to push and rotate catheters and wires to achieve this).

 

Months 10-15: Use the in vitro experimental set up and record*fluoroscopic motion (kinematics) of catheter and wire. Start with in silico model, apply transfer learning to in vitro data to finesse model.

 

Months 16-17: Use retrospective fluoroscopic in vivo routine thrombectomy navigation images from KCH to similarly finesse model.**

Months 18-24: In vitro operator-assist mode experiments: on approach to various anatomical structures information will be given in form of verbal (similar to car SatNav“rotate left in 5 mm”and/or visual information “e.g. arrow direction display”).

 

Months 12-25  Design responder robot (mimicking motion-capture gloves).Using:

*Additional collection of data using team’s motion-capture of gloves.

**Additional collect of in vivo routine thrombectomy data prospectively at KCH with motion-capture of gloves worn by the operator.

 

Months 25–42: Submit paper 1.

In vitro autonomous mode experiments: AI is controller and on approach to various anatomical structures information will be given to responder robot (mimicking motion-capture gloves).

 

Months 42-48: Submit paper 2.

Write up thesis.

  1. Hankey GJ. Stroke. Lancet. 2017;389:641-654.

  2. Saver et al. Time to Treatment With Endovascular Thrombectomy and OutcomesFrom Ischemic Stroke: A Meta-analysis. JAMA 2016 Sep 27;316(12):1279-88.

  3. Sentinel Stroke Audit Programme (SSNAP). Apr 2015–Mar 2016 –Annual Results Portfolio. National Results April 2015–Mar 2016, www.strokeaudit.org/ Documents/ Results/National/Apr2015Mar2016/Apr2015Mar2016 AnnualResults Portfolio.aspx(2016, accessed 1 January 2021).

  4. Flynn D et al. Intra-arterial mechanical thrombectomy stent retrievers and aspiration devices in the treatment of acute ischaemic stroke: a systematic review and metaanalysis with trial sequential analysis. Eur Stroke J 2017; 2: 308–318.

  5. Smith WSet al. Significance of large vessel intracranial occlusion causing acuteischemic stroke and TIA. Stroke 2009; 40: 3834–3840.

  6. Goyal M et al. Endovascular thrombectomy after large-vessel ischaemic stroke: a meta-analysis of individual patient data from five randomised trials. Lancet 2016; 387: 1723–1731.

  7. McMeekin P et al. Estimating the number of UK stroke patients eligible for endovascular thrombectomy: estimating the number of UK stroke patients eligiblefor endovascular thrombectomy. Eur Stroke J 2017; 2: 319–326.

  8. CrinnionW, JacksonB, SoodA,Lynch J, BergelesC, LiuH, RhodeK,Mendes PereiraV, Booth TC.Robotics in Neurointerventional Surgery: A Systematic Review of the Literature. JNIS In Press

  9. Vuong SM et al Applications of emerging technologiesto improve access to ischemicstroke care. Neurosurg Focus 2017 42 (4) E8

  10. Lu WS et al. Clinical application of a vascular interventional robot in cerebralangiography. Int J Med Robot 12:132–136, 2016

  11. Britz et al. Feasibility of Robotic-Assisted Neurovascular Interventions: Initial Experience in Flow Model and Porcine Model. Neurosurgery 2019 Apr 17. pii:nyz064. doi: 10.1093/neuros/nyz064. [Epub ahead of print]

  12. Lee MA et al. Making Sense of Vision and Touch: Learning Multimodal Representations for Contact-Rich Tasks. arXiv 2019 [Lee2019].

  13. Fang K et al. Discovering Generalizable Skills via Automated Generation of Diverse Tasks. arXiv 2021. [Kuan2021].

  14. Fagogenis G et al., Autonomous robotic intracardiac catheter navigation using haptic visionSci. Robot. 4, eaaw1977 (2019)

  15. Nahami K et al. HJB-RL: Initializing Reinforcement Learning with Optimal Control Policies Applied to Autonomous Drone Racing. Stanford.edu 2021[Nagami2021].

  16. Zachares PA et al. Interpreting Contact Interactions to Overcome Failure in Robot Assembly Tasks. arXiv 2021.[Zachares2021].

  17. ZengelerNet al.Hand Gesture Recognition in Automotive Human–Machine Interaction Using Depth Cameras. Sensors2019,19(1), 59.

  18. van Amstredam Bet al. Gesture Recognition in Robotic Surgery: a Review. IEEE Transactions on Biomedical Engineering, 2021 [vanAmsterdam2021

Saver et al. Time to Treatment With Endovascular Thrombectomy and Outcomes From Ischemic Stroke: A Meta-analysis. JAMA 2016 Sep 27;316(12):1279-88.

Ivanovic B https://ai.stanford.edu/blog/trajectory-forecasting/  2020 Accessed October 2021