💻 computer science

Learning-Based Modeling of a Magnetically Steerable Soft Suction Device for Endoscopic Endonasal Interventions

This paper presents a learning-based modeling framework for a miniaturized, magnetically steerable soft suction device used in endoscopic endonasal surgery, which utilizes embedded Fiber Bragg Grating sensors and a Random Forest algorithm to accurately predict the robot's 3D shape from magnetic actuation parameters with sub-millimeter error.

Original authors: Majid Roshanfar, Alex Zhang, Changyan He, Amir Hooshiar, Dale J. Podolsky, Thomas Looi, Eric Diller

Published 2026-02-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Majid Roshanfar, Alex Zhang, Changyan He, Amir Hooshiar, Dale J. Podolsky, Thomas Looi, Eric Diller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a surgeon trying to remove a brain tumor through a patient's nose. It's like trying to perform delicate surgery inside a narrow, winding cave using a rigid, straight stick. If the stick hits a sharp turn, it can't bend, and it might accidentally poke something fragile, like a nerve or a blood vessel.

To solve this, the researchers in this paper built a tiny, flexible, "smart" straw that can bend on command without any wires or motors inside it. Here is how they made it work, explained simply:

1. The "Magic Straw" (The Device)

Think of the device as a super-thin, flexible tube (about the width of a pencil lead) made of soft, medical-grade silicone.

The Magnet: At the very tip of this straw, they stuck a tiny, hollow magnet.
The "Remote Control": Instead of pushing a wire from the outside, they use a giant, invisible magnetic field generated by a table of coils underneath the patient. By spinning this magnetic field, they can make the magnet at the tip of the straw turn and pull the whole straw into a curve. It's like using a giant magnet to steer a tiny metal boat through a maze.
The "Eyes": Inside the straw, they ran a special fiber-optic cable (like a high-tech nerve) that acts as a ruler. As the straw bends, this cable stretches and compresses slightly. By measuring these tiny changes, the computer knows exactly what shape the straw is making in real-time.

2. The Problem: It's Too Complicated to Calculate

The researchers tried to use old-school math to predict how the straw would bend. But soft rubber is tricky. It doesn't bend like a stiff metal rod; it stretches, squishes, and reacts differently depending on how fast you spin the magnetic field. It's like trying to predict exactly how a wet noodle will flop around in the wind using a simple formula—it's too messy and unpredictable.

3. The Solution: Teaching a Computer to "See" the Shape

Instead of trying to write a perfect math equation for the rubber, the researchers decided to teach a computer by example.

The Training: They took the straw and spun the magnetic field in 70 different ways (changing the strength, speed, and distance). They recorded exactly what shape the straw made every single time. They collected over 5,000 examples of "Magnetic Field Input" $\rightarrow$ "Resulting Shape."
The Shortcut (Bezier Points): To make the data easier for the computer to learn, they didn't try to map every single millimeter of the straw. Instead, they imagined the straw's curve as a smooth line drawn by a digital pen. They only needed four invisible "control points" to define that curve. Think of these points like the handles on a curve tool in a drawing program; move the handles, and the whole line changes smoothly.
The Learning: They fed all 5,000 examples into two different types of computer brains (a Neural Network and a Random Forest) to see which one could best guess where those four control points would be based on the magnetic field settings.

4. The Results: The "Forest" Won

The computer learned the pattern incredibly well.

The Winner: One type of computer model (called a Random Forest) was much better than the other. It could predict the shape of the straw with sub-millimeter accuracy (less than the thickness of a human hair).
Why it matters: The model learned that the strength of the magnetic field mostly controls how far the tip bends, while the speed of the spin affects how the base of the straw moves.
Speed: The computer can make this prediction almost instantly, which is fast enough to be used while a surgeon is actually moving the tool.

5. What This Means (According to the Paper)

The paper claims this is the first time a learning-based model has been used specifically for a magnetically steered soft suction tool with these specific sensors.

No More Guessing: It removes the need for complex physics equations that assume the rubber is perfectly uniform (which it never is).
Real-Time Feedback: Because the model is fast and accurate, it can tell the surgeon exactly where the tip of the tool is, even if the tool is twisting and turning deep inside the nose.
Safety: By knowing the shape precisely, surgeons can navigate tight, dangerous spaces without accidentally hitting critical brain structures.

In short: The researchers built a tiny, bendy straw that can be steered by invisible magnets. They taught a computer to predict exactly how that straw will bend by showing it thousands of examples, creating a "smart" system that can guide surgeons through the narrow, winding tunnels of the human nose with extreme precision.

Technical Summary: Learning-Based Modeling of a Magnetically Steerable Soft Suction Device

Problem Statement
Endoscopic endonasal brain tumor resection is a minimally invasive technique that relies on navigating narrow anatomical corridors. However, the maneuverability of conventional rigid instruments is often limited by patient-specific anatomical complexities (e.g., narrow nasal passages, reduced skull base), increasing the risk of damage to critical neurovascular structures. While flexible robotic instruments offer a solution, their inherent nonlinear, hyperelastic behavior and lack of intuitive control pose significant challenges. Traditional physics-based modeling approaches (e.g., Constant Curvature, Cosserat rod theory, Geometric Variable Strain) struggle with magnetically actuated soft devices due to the difficulty in precisely characterizing 3D-printed material properties, the complexity of configuration-dependent magnetic torques, and the computational cost of real-time simulation. Furthermore, existing learning-based studies have primarily focused on tendon-driven or pneumatic robots, leaving a gap in data-driven modeling for magnetically actuated soft surgical tools with embedded sensing.

Methodology
The authors propose a data-driven modeling framework that maps magnetic actuation inputs directly to the geometric configuration of a custom soft suction device.

Device Design: A miniaturized (4 mm outer diameter, 2 mm inner channel, 40 mm active length) soft suction device was 3D printed using biocompatible SIL 30 silicone. It features a distal permanent magnet for steering and an embedded multi-core Fiber Bragg Grating (FBG) sensor for real-time shape sensing.
Shape Encoding: Raw FBG strain data is processed to reconstruct the 3D centerline of the robot. To create a compact, low-dimensional representation suitable for learning, the reconstructed shape is fitted to a cubic Bezier curve defined by four control points ( $p_0$ to $p_3$ ). The endpoints ( $p_0, p_3$ ) are constrained to the device base and tip, while intermediate points ( $p_1, p_2$ ) define the curvature profile.
Experimental Data Collection: Experiments were conducted using an 8-coil electromagnetic system. The device was actuated under rotating magnetic fields with varying parameters: magnitude (0–14 mT), frequency (0.2–1.0 Hz), and vertical tip distance from the coil table (90 mm and 100 mm). This resulted in a dataset of 5,097 experimental samples.
Modeling Approaches: Two machine learning architectures were trained to predict the four Bezier control points from the magnetic field parameters ( $b_x, b_y$ $b_{x}, b_{y}$ , magnitude, frequency, and distance):
1. Neural Network (NN): A feedforward network with three hidden layers.
2. Random Forest (RF): An ensemble of 12 regressors, each with 200 decision trees.

Key Contributions

Device Fabrication: The design and fabrication of a miniaturized, 3D-printed, biocompatible soft suction device with integrated FBG sensing, specifically tailored for endoscopic endonasal neurosurgery.
Novel Modeling Paradigm: The introduction of a learning-based framework that directly links magnetic field inputs to Bezier control points. This approach bypasses the need for explicit constitutive parameter identification and analytical kinematic derivations, offering a compact and interpretable geometric representation.
Unified Framework: The conceptual integration of magnetic field characterization, embedded FBG sensing, and Bezier-based learning into a single modeling strategy applicable to other magnetically actuated continuum robots.
Experimental Validation: Comprehensive validation demonstrating that the proposed framework can accurately model the complex, nonlinear, and frequency-dependent behavior of hyperelastic soft robots under magnetic actuation.

Results

Model Performance: The Random Forest (RF) model significantly outperformed the Neural Network (NN) across all metrics. The RF model achieved a mean Root Mean Square Error (RMSE) of 0.087 mm in predicting the control points, compared to 0.121 mm for the NN.
Shape Reconstruction: The RF model yielded a mean shape reconstruction error of 0.064 mm. The error distribution was tighter and more centered around zero for the RF model, particularly for the distal control points ( $p_2$ and $p_3$ ), which are critical for tip positioning.
Feature Importance: Analysis revealed that magnetic field components ( $b_x, b_y$ ) predominantly influence the distal control points, while actuation frequency and vertical distance significantly affect the base configuration and vertical positioning.
Computational Efficiency: The RF model demonstrated high computational efficiency suitable for real-time inference, with a prediction time per sample of approximately $7.89 \times 10^{-5}$ seconds.

Significance and Claims
The paper claims to present the first application of a learning-based shape prediction model specifically tailored for magnetically actuated soft surgical devices equipped with embedded FBG sensing. The significance of this work lies in its ability to model the complex nonlinear behavior of soft robots without relying on simplified physical assumptions or difficult-to-obtain material parameters. By achieving sub-millimeter prediction accuracy and real-time inference, the framework establishes a foundation for the intelligent control of magnetically actuated tools in minimally invasive neurosurgery.

The authors emphasize that while the learning-based approach offers a pragmatic pathway for rapid device characterization, it is complementary to physics-based frameworks. They acknowledge limitations, including the reliance on a single prototype (device-specific), the use of idealized rotating magnetic fields, and the absence of tissue interaction effects in the current training data. Consequently, the authors position this work as a necessary baseline characterization, with future work required to validate generalization across multiple devices, incorporate tissue interaction data, and test performance in complex clinical magnetic environments.