The Feasibility of Haar Feature-Based Endoscopic Ultrasound Probe Tracking for Implanting Hydrogel Spacer in Radiation Therapy for Pancreatic Cancer

Ziwei Feng, Hamed Hooshangnejad, Eun Ji Shin, Amol Narang, Muyinatu A. Lediju Bell, Kai DingFrontiers in Oncology 2021 · DOI

TL;DR

This paper introduces a novel Haar feature-based method for tracking endoscopic ultrasound (EUS) probe positions without external hardware. This technique aims to improve the accuracy and efficiency of implanting hydrogel spacers for pancreatic cancer radiation therapy by aligning real-time EUS images with pre-acquired CT/MRI scans.

What Problem Does This Paper Address?

Pancreatic cancer is a deadly disease, and improving radiation therapy outcomes often involves dose escalation. However, delivering higher doses is challenging due to the proximity of radiosensitive organs like the duodenum. Hydrogel spacers can create a buffer, reducing radiation dose to these organs. While EUS-guided hydrogel injection shows promise, accurately placing the spacer is difficult due to the uncertainty of optimal injection points and the dynamic nature of the anatomy. Current tracking methods rely on external hardware, which can be cumbersome. This paper addresses the need for a robust, hardware-free method to precisely track the EUS probe's position relative to pre-treatment imaging (CT/MRI) to guide hydrogel injection accurately.

Methods

The core of the proposed method is a Haar feature-based approach for image registration. The process involves:

  1. Simulated EUS Image Generation: Using the Field II simulation software, realistic EUS images are generated based on anatomical information from CT or MRI scans. This creates a "simulated EUS image dictionary."
  2. Haar Feature Extraction: Efficient Haar features are extracted from these simulated EUS images. These features are known for their ability to represent image information with high accuracy and speed, making them suitable for object detection and image registration.
  3. Probe Position Estimation: The relative position of the EUS probe is estimated by comparing the Haar feature vector of a real-time EUS image (the "target EUS image") with the Haar feature vectors in the dictionary. The best match is found using normalized cross-correlation (NCC).
  4. Validation: The method's feasibility is tested using both a phantom experiment and real patient data (CT and MRI scans).

The novelty lies in adapting Haar features for this specific EUS probe tracking task, enabling a direct link between 2D EUS views and 3D pre-diagnostic scans without external tracking devices.

Datasets

  • Simulated EUS Image Dictionary: Generated using the Field II simulation software based on anatomical data from CT and MRI scans. For phantom experiments, this dictionary contained 123 simulated images. For patient experiments, the size and composition varied.
  • Phantom Data: An EUS training phantom (ATS Laboratories, Model GIETP) with four echogenic sphere targets was used. This phantom was scanned using a CT simulator.
  • Patient Data:
    • CT Scans: Obtained from patients who underwent hydrogel injection. These were interpolated to create ICT (Interpolated Computed Tomography) datasets.
  • MRI Scans: Obtained from patients before hydrogel injection (O-MRI) and simulated post-injection MR images (S-MRI) were generated.

Key Insights and Ideas

  • Hardware-Free Tracking: The most significant insight is the successful demonstration of EUS probe tracking without relying on external hardware, which simplifies the procedure and potentially reduces costs.
  • Haar Features for EUS: Haar features, traditionally used in face detection and image compression, prove to be effective for capturing the subtle anatomical variations in EUS images, enabling robust registration.
  • Bridging 2D and 3D: The method effectively bridges the gap between real-time 2D EUS imaging and pre-acquired 3D CT/MRI data, which is crucial for accurate spatial localization.
  • Phantom vs. Real-World Performance: While the method performs well on a simplified phantom with spherical targets, its performance in more complex patient anatomy (especially with lower MRI resolution) highlights the challenges of real-world application.

Key Results

  • Phantom Experiments: The method successfully found the best matched simulated EUS image from a dictionary of 123 images. The errors for four target points between the real EUS image and the best matched simulated image were within 1 mm.
  • Patient CT/MRI Scans: The method accurately identified the best matched simulated EUS image, confirming probe location. However, performance in MRI images was less robust due to lower image resolution.
  • Matching Accuracy: Analysis showed that increasing the scale parameter of Haar features improved matching accuracy, especially for lower-quality simulated images with less defined edges. The number of basic Haar feature functions showed less significant impact on matching error.

Limitations and Future Work

  • Image Resolution: The method's robustness is affected by image resolution, particularly in MRI scans, leading to less accurate tracking.
  • Simulation Complexity: Generating a comprehensive simulated EUS dictionary with fine intervals and varied probe directions is computationally intensive and time-consuming.
  • Limited Ground Truth: In patient experiments, the lack of precise ground truth for probe location made direct error quantification challenging.
  • Anatomical Complexity: The current phantom is simplified; more complex phantoms are needed to mimic realistic endoscopic injection scenarios.
  • Motion Artifacts: Breathing motion, a significant factor in real-time EUS, was not explicitly modeled in the simulation, which could impact accuracy.

Future Work Suggestions:

  • Improve Simulation: Refine simulated EUS images with finer intervals and more varied probe directions.
  • Enhance Image Quality: Develop pre-processing steps to improve the quality of both simulated and real EUS images.
  • Incorporate Motion Models: Integrate models for breathing motion to improve robustness.
  • External Tracking Validation: Consider using external tracking hardware (e.g., infrared markers) to establish ground truth for more rigorous evaluation.
  • Combine with Simulation Models: Integrate the tracking method with physics-based simulation models (like FEMOSSA) for pre-treatment planning and guidance.

Summary updated

Abstract

Our Haar feature-based method is capable to find the best matched simulated EUS image from the dictionary. We demonstrated the feasibility of our method for tracking EUS probe without external tracking hardware, thereby guiding the hydrogel injection between the head of the pancreas and duodenum.
BibTeX
@inproceedings{feng2021feasibility,
  author = {Ziwei Feng and Hamed Hooshangnejad and Eun Ji Shin and Amol Narang and Muyinatu A. Lediju Bell and Kai Ding},
  title = {{The Feasibility of Haar Feature-Based Endoscopic Ultrasound Probe Tracking for Implanting Hydrogel Spacer in Radiation Therapy for Pancreatic Cancer}},
  booktitle = {Frontiers in Oncology},
  year = {2021},
  doi = {10.3389/fonc.2021.759811},
}
Figure