⚡ electrical engineering

A Dual-Modal Fiducial Marker with Single-View X-ray Pose Recoverability for MR-Guided Surgical Navigation

This study proposes and validates a novel dual-modal fiducial marker featuring a central cube and four cylindrical axes that satisfies the specific constraints of MR-guided surgical navigation by enabling single-view X-ray pose recoverability and achieving high-accuracy optical recognition.

Original authors: Ganzhe Xu, Xi Li, Wenyi Xiong, Zhen Yin, Yaxin Yang, Xiaopeng Wang, Da Zhong, Hua Liu, Yang Xie, Kunli Chen, Chenggong Wang

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ganzhe Xu, Xi Li, Wenyi Xiong, Zhen Yin, Yaxin Yang, Xiaopeng Wang, Da Zhong, Hua Liu, Yang Xie, Kunli Chen, Chenggong Wang

Original paper licensed under CC BY 4.0 (https://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 you are a surgeon trying to perform a delicate operation. You need a high-tech "GPS" to guide your tools, but there's a catch: the GPS camera is strapped to your head (like a pair of smart glasses), and it moves every time you turn your head.

The Problem:
Old-school surgical markers are like lighthouses designed for a stationary observer standing on a hill. They work great if the camera is fixed in one spot, but if the camera moves around you (like a surgeon's head), the marker might get hidden behind a hand or a tool, or the camera might look at it from an angle where it looks like a flat, useless blob. Also, surgeons often need to check X-rays during surgery, but standard markers are invisible or confusing on X-rays.

The Solution: The "Magic Dice"
The researchers created a new kind of marker that solves these problems. Think of it as a specialized, multi-sided die with four long, uneven "legs" sticking out of it.

The Cube (The Face): The center is a cube. Just like a die has six faces, this cube has corners and edges that are easy for the "smart glasses" (the camera on your head) to see from almost any angle. Even if you look at it from the side or top, the camera can figure out exactly where it is.
The Legs (The X-Ray Clues): Sticking out of the cube are four cylindrical "legs" (axes). Here is the clever part: each leg is a different thickness.
- Imagine looking at a group of pencils of different sizes through a shadow (an X-ray). Even if you only see their shadows, you can tell which pencil is which because the shadows have different widths.
- Because the legs are different sizes, the computer can look at a single X-ray image, measure the width of the shadows, and instantly know exactly how the marker is rotated and where it is in 3D space.

How It Works (The "Magic Trick"):
The paper claims this marker is a "dual-modal" superhero. It speaks two languages at once:

Language 1 (Optical): The cube's shape lets the smart glasses see it clearly, no matter how the surgeon moves.
Language 2 (X-Ray): The different-sized legs let the X-ray machine see it clearly, even though X-rays usually just show flat shadows.

The Results:
The team tested this "Magic Dice" 100 times with X-rays and nearly 200 times with optical cameras.

X-Ray Performance: When looking at the X-ray shadows, the computer could figure out the marker's angle with an average error of about 1.5 to 1.9 degrees. That's roughly the width of a pencil held at arm's length. It worked even when the marker was tilted at extreme angles, though it got slightly less accurate if the marker was turned sideways so much that one leg was hidden.
Optical Performance: When the "smart glasses" looked at the marker, it was even more precise, with an average error of about 1.5 degrees. It successfully found the marker in over 90% of the attempts.
The "Transfer" Test: They also proved that if you know where the marker is, you can mathematically "transfer" that knowledge to a nearby object (like a block of wood). The computer could draw a wireframe over the wood that matched the real wood very closely, proving the system can link the marker's position to the patient's anatomy.

The Big Takeaway:
This paper doesn't claim to have fixed all surgery yet, but it proves that this specific "Magic Dice" design works. It is the first marker that can be seen clearly by a moving head-mounted camera and by an X-ray machine at the same time, using a single, simple object. It solves the "blind spot" problem of old markers by being round and multi-sided, and it solves the "invisible on X-ray" problem by using legs of different sizes.

Technical Summary: A Dual-Modal Fiducial Marker with Single-View X-ray Pose Recoverability for MR-Guided Surgical Navigation

1. Problem Statement

Traditional surgical navigation systems rely on fixed external reference frames and infrared reflective spheres. While effective in conventional setups, these markers are unsuitable for Mixed Reality (MR) surgical navigation using head-mounted displays (HMDs). In MR navigation, the tracking camera moves with the surgeon's head, creating a dynamic reference frame where the line of sight overlaps with the surgical field. This introduces three critical challenges that existing markers cannot simultaneously address:

Omnidirectional Recognizability: Conventional markers (e.g., spherical retroreflectors or planar QR codes) suffer from optical blind spots or fail when viewed from the side/rear, which is problematic when multiple observers or instruments occlude the view.
X-ray Recognizability: Intraoperative updates of the marker-to-bone pose often rely on C-arm fluoroscopy. Existing markers (like planar codes or spheres) lack distinct geometric features in X-ray projections, preventing accurate pose estimation from a single view.
Multimodal Registration: There is a lack of markers that provide consistent, distinguishable geometric primitives (points, lines, planes) that can be stably extracted in both optical and X-ray modalities to enable robust cross-modal correspondence.

Current attempts to adapt conventional workflows to HMDs result in sharp increases in registration error, primarily because the physical constraints of the HMD (moving camera, overlapping field of view) were not considered in the marker design.

2. Methodology

The authors propose a novel fiducial marker design and a corresponding pose estimation framework based on three derived design constraints.

A. Marker Design
The proposed marker consists of a central cube and four cylindrical axes extending from the cube's center toward the vertices of a regular tetrahedron.

Central Cube: Provides rich optical features (corners and faces) for omnidirectional recognition by the HMD camera.
Four Cylinders: These have unequal diameters ( $r_1 < r_2 < r_3 < r_4$ ). This specific geometry allows the cylinders to be distinguished by their projected widths in X-ray images.
Rigid Connection: The cube and cylinders are rigidly connected, ensuring a fixed spatial transformation between the optical and X-ray poses.

B. Theoretical Foundation
The study mathematically proves that the marker's 6 Degrees of Freedom (6-DoF) pose can be uniquely determined from a single X-ray image under a known projection system.

Geometric Constraints: Each cylinder provides three independent measurements (centerline direction and projected width). Four cylinders yield 12 measurements to solve for 6 unknowns (3 rotation, 3 translation), creating an overdetermined system.
Ambiguity Resolution: The unequal radii resolve reflection ambiguities (mirror solutions) and provide absolute scale, eliminating the "object scaling + distance scaling" ambiguity common in single-view pose estimation.
Robustness: The tetrahedral arrangement ensures that even if one cylinder is parallel to the X-ray beam (degenerating its projection), the remaining three non-coplanar cylinders are sufficient to determine the rotation matrix.

C. Algorithm and Data

Pose Estimation: A cascaded machine learning framework is employed.
- X-ray Branch: Uses a ResNet50/HopeNet network to estimate rotational priors, combined with geometric constraints to recover translation.
- Optical Branch: Uses YOLOv11 for bounding box detection and U-Net for keypoint prediction (cube corners/cylinder tips), followed by a Levenberg-Marquardt optimizer to minimize reprojection error.
Data Acquisition:
- X-ray: 100 valid samples acquired using a GE Definium 6000 system with an IMU-based ground truth reference.
- Optical: 194 valid samples acquired using a Canon R7 camera simulating HMD perspectives, with checkerboard references for ground truth.

3. Key Results

The study validated the marker's performance on 100 X-ray sets and 194 optical sets.

A. X-ray Pose Recognition

Accuracy: The mean absolute error (MAE) was 1.584° ± 1.256° in the coronal plane (X-axis) and 1.898° ± 1.502° in the sagittal plane (Y-axis).
Composite Error: The overall angular error had a median of 2.45° and an MAE of 2.73°.
Distribution: Errors were symmetrically zero-centered. The system remained solvable even under extreme viewing angles, though errors increased slightly when a cylinder aligned with the beam (degraded but solvable configuration).

B. Optical Pose Recognition

Success Rate: The detection success rate was 90.2% (175/194 valid comparisons).
Accuracy: Among reliable samples, the MAE was 1.521° ± 0.976°.
- Yaw: 0.252° (highest accuracy).
- Pitch: 0.411°.
- Roll: 1.438° (lowest accuracy, consistent with weaker constraints on in-plane rotation).
Failure Modes: 4 cases (2.3%) showed complete pose loss (error > 170°) under extreme viewing angles, attributed to reflective ambiguity.

C. Pose Transfer and Cross-Modal Consistency

Transferability: A proof-of-concept experiment transferring the marker pose to a target object (wooden block) demonstrated that while point deviations increased with depth (26.6 px mean), edge angular deviations remained low (3.10° mean) and uncorrelated with point errors. This indicates the pose estimation preserves structural orientation fidelity.
Consistency: The optical modality showed higher precision (median error 1.38°) compared to X-ray (2.45°), but both modalities exhibited similar error patterns: axes aligned with dominant geometric features were most accurate, while in-plane rotational axes were least accurate.

4. Significance and Claims

The paper claims to be the first study to systematically derive marker design requirements from the physical constraints of HMD-based navigation rather than adapting existing markers.

Theoretical Contribution: It establishes three first-principles constraints (omnidirectional recognizability, X-ray recognizability, and multimodal registration) that define the requirements for MR surgical navigation markers.
Design Innovation: It demonstrates that a simple geometric structure (cube + unequal cylinders) can simultaneously satisfy these constraints, enabling single-view X-ray pose recoverability and robust optical tracking.
Practical Impact: The results indicate that this design methodology serves as a viable reference for developing MR surgical navigation markers, potentially reducing registration errors caused by the "attention shift" and dynamic reference frames inherent in HMD systems.

The authors maintain a modest tone, noting that while the marker's dual-modal recognition is validated, the full pipeline including bone recognition and relative pose transfer in complex surgical scenarios remains a subject for future work. The current study focuses strictly on validating the feasibility of the marker's recognition capabilities under both imaging modalities.

Technical Summary: A Dual-Modal Fiducial Marker with Single-View X-ray Pose Recoverability for MR-Guided Surgical Navigation

1. Problem Statement

2. Methodology

3. Key Results

4. Significance and Claims

More like this