Rapid Path Planning Algorithm for Percutaneous Rigid Needle Biopsy Based on Optical Illumination Principles

Jian Liu, Kang Shuai, Juan Ren, Dongxia Zhang, Bing Niu, Kai XuSensors 2025 · DOI

TL;DR

This paper introduces a novel, ultra-fast algorithm for planning percutaneous rigid needle biopsy paths by reframing the problem as an optical illumination simulation. By leveraging GPU acceleration and a simplified vascular model, the method achieves near-real-time planning (under 2 seconds) and demonstrates high clinical feasibility, significantly improving safety and efficiency in minimally invasive procedures.

What Problem Does This Paper Address?

Percutaneous needle biopsies are crucial for diagnosing conditions like cancer, but selecting the optimal needle trajectory is challenging. This is because:

  • Time Sensitivity: Planning relies on CT scans, and delays increase the risk of patient movement invalidating the plan.
  • Anatomical Complexity: Navigating around vital structures like blood vessels, bones, and organs is critical but computationally intensive.
  • Traditional Methods are Slow: Existing methods, whether manual or algorithmic, can take seconds to minutes, which is too long in time-sensitive procedures.
  • Vascular Modeling Challenges: Accurately modeling blood vessels without excessive computational cost is difficult, often leading to overly conservative "forbidden zone" approaches that limit viable paths.

This paper aims to address these issues by developing a rapid, accurate, and clinically relevant path planning system.

Methods

The core of the proposed method lies in an "optical illumination-based" approach:

  • 3D Reconstruction and Segmentation: CT images are used to create a 3D model of anatomical structures. A segmentation model (SegResNet) is employed to identify organs, tissues, and lesions with high accuracy (average DSC of 0.9122).
  • Optical Simulation: The target lesion is treated as a virtual light source. Tissues are assigned "absorption coefficients" based on their risk (e.g., opaque for bones, partially translucent for vessels).
  • Light Propagation: Simulating light propagation models needle trajectory.
  • Diffuse Reflection: Simulates needle angle constraints.
  • Accumulated Absorption: Represents tissue depth and puncture constraints.
  • Grayscale Map Generation: This simulation results in a grayscale map on the skin surface, highlighting potential puncture sites based on "feasibility intensity."
  • Machine Learning for Preference Modeling: A random forest model is trained to mimic a surgeon's preferences for puncture sites. It analyzes an RGB image derived from the grayscale map to automatically select the optimal path and entry point.
  • GPU Acceleration: The entire process is heavily optimized using GPU acceleration and techniques like early ray termination and empty space skipping to achieve ultra-fast computation.
  • Simplified Vascular Modeling: Instead of explicit segmentation, vessels are assigned absorption coefficients, allowing for a more nuanced risk assessment and avoiding computationally expensive detailed modeling.

Datasets

  • LUNA16 Dataset: Used for training and validating the segmentation models. This dataset contains 888 CT scans with annotations for lung nodules.
  • Internal Clinical Dataset: Used for testing the robustness and generalization of the segmentation model.
  • Tianchi Lung Nodule Dataset and LNDb Dataset: Publicly available datasets used for robustness testing of the segmentation model.
  • Annotated Projected Images (250 cases): Used for training and evaluating the machine learning-based path recommendation. These images were derived from the LUNA16 dataset and annotated by a thoracic surgeon.

Key Insights and Ideas

  • Optical Analogy for Path Planning: Reframing needle path planning as a light propagation problem is a novel and computationally efficient approach. The straight-line nature of light mirrors rigid needle trajectories, and the physics of light absorption and reflection can effectively model anatomical constraints.
  • GPU-Accelerated Simulation: Leveraging modern GPU capabilities is crucial for making complex simulations computationally feasible in near real-time.
  • Vascular Risk as Absorption: Assigning absorption coefficients to vessels, rather than treating them as strict exclusion zones, allows for a more flexible and clinically relevant approach to vascular navigation. This acknowledges that some small vessels can be safely traversed.
  • Machine Learning for Surgeon Preference: Using a random forest to learn and replicate a surgeon's decision-making process for selecting optimal entry points is a practical way to incorporate expert knowledge into an automated system.
  • Grayscale Map as Intuitive Visual Aid: The generated grayscale map provides a visually intuitive representation of puncture site desirability, aiding rapid assessment and potentially simplifying 3D anatomy interpretation.

Key Results

  • Ultra-Fast Computation: The entire path planning pipeline (segmentation to recommendation) achieves an average runtime of 1.905 seconds per sample, a significant speedup compared to traditional methods (seconds to minutes).
  • High Segmentation Accuracy: The SegResNet model achieved an average Dice Similarity Coefficient (DSC) of 0.9122 for anatomical structures.
  • Clinical Feasibility: A thoracic surgeon assessed 50 recommended paths and found:
  • 78% rated as fully satisfactory.
  • 22% rated as moderately satisfactory.
  • 0% deemed unsatisfactory.
  • 100% considered clinically feasible.
  • Machine Learning Performance: The random forest model achieved a high AUC-ROC of 0.948 for predicting surgeon preferences, with good precision (0.706) and recall (0.766).
  • Robustness: The segmentation model showed good generalization across different datasets, achieving an average DSC of 0.8563 on a combined set of public and private clinical data.

Limitations and Future Work

  • Limitations:
    • Lung Nodule Segmentation: Performance was weaker for lung nodules compared to other organs due to their small size, ambiguous boundaries, and limitations in the LUNA16 dataset's annotation format (spherical representation). This resulted in higher ASD and HD95 metrics for nodules.
  • Data-Driven Model Dependence: The machine learning model's performance relies on the quality and representativeness of the training data regarding surgeon preferences.
  • Future Work:
    • Improved Segmentation Models: Developing more accurate and robust segmentation models for target organs and lung nodules is crucial.
  • Real-World Application and Validation: Translating the method into practical clinical settings with user-friendly interfaces and integration into existing systems.
  • In Vivo Studies: Conducting comprehensive validation on diverse patient datasets and ultimately in vivo studies to evaluate clinical performance, safety, and efficiency.
  • Exploring Pareto Optimization: Investigating the integration of Pareto optimization with the optical model to potentially offer both speed and explicit multi-objective control.

Summary updated

Abstract

Optimal needle trajectory selection is critical in biopsy procedures to minimize tissue damage and ensure diagnostic accuracy. Timely trajectory planning is essential, as it relies on preoperative CT imaging. Prolonged processing times increase the risk of patient movement, rendering the planned path invalid. Traditional methods relying on clinician expertise or slow algorithms struggle with complex anatomical modeling for structures such as blood vessels. We introduce a novel method that reframes trajectory planning as an optimal puncture site identification problem by leveraging optical principles and computer rendering. A 3D model of key anatomical structures is reconstructed from CT images and segmented using SegResNet (average Dice similarity coefficient of 0.9122). A virtual light source positioned at the target illuminates the space, assigning distinct absorption coefficients to tissues based on needle permissibility and risk. Diffuse reflection simulates needle angle, and accumulated absorption represents depth, capturing puncture constraints. This simulation generates a grayscale map on the skin surface, highlighting candidate puncture sites. Furthermore, we employ a random forest-based method to model clinician preferences. This model analyzes an RGB image derived from the grayscale distribution to automatically select the optimal path and determine the needle entry point. The experimental evaluation demonstrates an average computation time of just 1.905 s per sample, which is significantly faster than traditional methods that require seconds to minutes. Moreover, clinical assessment by a thoracic surgeon found that 78% of the recommended paths met clinical standards, with 0% deemed unsatisfactory. These findings suggest that our method provides a rapid, intuitive, and reliable decision-support tool, improving biopsy safety and efficiency.
BibTeX
@inproceedings{liu2025rapid,
  author = {Jian Liu and Kang Shuai and Juan Ren and Dongxia Zhang and Bing Niu and Kai Xu},
  title = {{Rapid Path Planning Algorithm for Percutaneous Rigid Needle Biopsy Based on Optical Illumination Principles}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25072137},
}
Figure