Abstract
The purpose of this study is to design a 3D navigation strategy with 2D ultrasonic images, based on the assumption that internal target trajectories can be evaluated using external surrogate signals. This paper proposes a 3D navigation strategy and a fast tumor tracking system utilizing learning-based methods to map the relationship between external surrogate signals and internal tumor trajectories. Using retrospective clinical data, the experimental results demonstrate that the system achieves high-accuracy tumor tracking and navigation, suitable for clinical applications.
Introduction
In interventional ablation surgeries, such as radiofrequency or microwave ablation, precision is paramount. However, respiratory motion causes significant organ displacement; for instance, the liver can move between 10mm and 55mm during respiration. This motion often forces surgeons to expand treatment margins, leading to the destruction of healthy tissue. While real-time imaging like fluoroscopy can track tumors, it involves ionizing radiation. Ultrasound (US) is a safer alternative but is typically limited to 2D views, making 3D navigation difficult. This research presents a method to predict 3D internal tumor positions using non-invasive, radiation-free external signals.
Methodology
The authors developed a multi-component navigation strategy combining electromagnetic tracking with machine learning.
Hardware Integration and Data Acquisition
The system utilizes the Aurora Electromagnetic (EM) Tracking System. Unlike optical systems, EM tracking does not require a clear line-of-sight, allowing surgeons to move freely without obstructing sensors. Seven magnetic sensors are placed on the patient’s chest and abdomen, generating 49 motion features. A 2D ultrasound probe is also fitted with an EM sensor to track its orientation and position.
Coordinate Calibration and Learning Model
A core component of the system is the coordinate transformation that maps 2D pixels from the ultrasound image into a 3D "world coordinate" system. This allows the system to localize internal targets in three dimensions.
The prediction engine uses a Gradient Boosting Machine (GBM), a supervised learning algorithm that models the complex, non-linear relationship between external skin movement and internal organ motion. To provide "ground truth" for training, a template-matching algorithm tracks stable landmarks (such as kidney structures) in the ultrasound video feed.
Key Results and Insights
The system was tested using retrospective clinical data, yielding several significant findings:
- High Precision: The system achieved a maximum estimation error of less than 1.9 mm when trained on data from the preceding 15 seconds.
- Temporal Accuracy: Accuracy decreases as the time from the training window increases, indicating that the model requires periodic updates to maintain precision.
- Non-Linear Modeling: Respiratory motion is non-linear and complex. The GBM handles these "weak linear" relationships more effectively than traditional linear models or Kalman filters.
- Computational Efficiency: The GBM approach provides high accuracy with lower computing time than complex deep learning alternatives, making it ideal for real-time surgical use.
Conclusion and Future Work
The study successfully demonstrates a real-time 3D tumor tracking system with sub-2mm accuracy. While promising, the authors note that future work should involve larger clinical datasets to ensure robustness across different body types. Additionally, optimizing feature selection could reduce the number of required sensors, and establishing a formal "period of validity" for the model would help surgeons know exactly when recalibration is necessary.
Summary updated