Robust fluoroscopic respiratory gating for lung cancer radiotherapy without implanted fiducial markers

Ying Cui, Jennifer Dy, G Sharp, Brian M. Alexander, Steve JiangPhysics 2007 · DOI

TL;DR

This paper presents a novel fluoroscopic-based respiratory gating technique for lung cancer radiotherapy that avoids the need for implanted markers. By comparing real-time X-ray images to pre-established templates of the tumor's position during exhalation, the system accurately triggers radiation delivery, significantly improving upon previous methods by employing template clustering for enhanced robustness and efficiency.

What Problem Does This Paper Address?

Lung cancer tumors often move significantly with respiration, posing a challenge for precise radiotherapy. Traditional methods to account for this motion include:

  • External surrogates: Using markers on the patient's body (e.g., abdomen) to infer tumor position. However, the correlation between external markers and internal tumor motion can be unreliable.
  • Implanted fiducial markers: Placing small markers directly near the tumor. While more accurate, this invasive procedure carries risks like pneumothorax and is not widely accepted.

This paper aims to overcome these limitations by developing a non-invasive, robust, and accurate respiratory gating system that relies solely on fluoroscopic (real-time X-ray) imaging. The goal is to ensure radiation is delivered only when the tumor is in the optimal treatment position, minimizing damage to surrounding healthy tissues.

Methods

The core of the proposed technique is a correlation score-based template matching approach. Here's a breakdown of their methods:

  • Training Phase (Patient Setup):
    • Fluoroscopic Image Acquisition: During patient setup, a sequence of orthogonal fluoroscopic images is captured.
  • Motion Enhancement: A pre-processing step enhances the visibility of the moving tumor by subtracting the average image from each individual frame. This highlights the dynamic structures.
  • Reference Template Generation: The key innovation lies in how these reference templates are created. They explore three methods:
  • Single Template Method: A single composite template is generated by averaging all motion-enhanced images acquired during the end-of-exhale (EOE) phase (identified by intensity thresholds).
  • Multiple Template Method: Each EOE image is treated as a separate template, and correlation scores are computed against all of them. These scores are then combined (e.g., averaged) to produce a final score.
  • Template Clustering Method: This is the most advanced approach. It uses clustering algorithms (specifically Gaussian Mixture Models with Principal Component Analysis for dimensionality reduction) to group similar EOE templates. This reduces a large number of potentially redundant templates into a few representative ones, which are then used for matching.
  • Treatment Phase:
    • Real-time Image Matching: Incoming fluoroscopic images during treatment are motion-enhanced.
  • Correlation Score Calculation: The motion-enhanced images are compared to the generated reference templates using a normalized correlation coefficient.
  • Gating Signal Generation: A threshold is applied to the correlation score. If the score exceeds the threshold (indicating the tumor is in the desired position), a gating signal is sent to turn on the radiation beam; otherwise, it remains off.

Datasets

  • Training and Evaluation Data: The study utilized fluoroscopic sequences from five lung cancer patients.
  • Data Characteristics: Each sequence was approximately 30 seconds long, acquired at 10 frames per second. A portion of this data (8-10 seconds, representing 2-3 breathing cycles) was used for training (template generation), and the rest for simulating treatment delivery.
  • Reference Gating Signals: For evaluation, reference gating signals were manually determined by a radiation oncologist who tracked the tumor's position across image frames.

Key Insights and Ideas

  • EOE Phase is Key: The end-of-exhale (EOE) phase is identified as the most stable and suitable phase for gating due to its longer duration and consistent tumor position.
  • Motion Enhancement is Crucial: Simply using raw fluoroscopic images is insufficient. The motion enhancement technique effectively isolates the moving tumor, making template matching more reliable.
  • Single Template Limitations: The single template method, while simple, proved to be highly sensitive to noise, leading to erratic gating signals.
  • Aggregation Improves Robustness: Both the multiple template and clustering methods showed improved robustness due to the "smoothing effect" of aggregating correlation scores from multiple templates. This mitigates the impact of noise from any single template.
  • Clustering Offers the Best Balance: The template clustering method emerged as the most promising, offering a good balance between computational efficiency and accuracy by reducing redundant templates while retaining essential information.
  • Error Location Matters: When errors did occur with the clustering method, they were typically near the edges of the gating window, suggesting that a slightly larger margin could further reduce dosimetric errors.

Key Results

  • Clustering Outperforms: The template clustering method consistently achieved the lowest false positive (FP), false negative (FN), and total error rates across all patients and duty cycles (35% and 50%).
  • Accuracy: The clustering method achieved accuracies ranging from 91% to 98.2% for the 35% duty cycle and 91% to 96.5% for the 50% duty cycle.
  • Reduced Noise: Gating signals generated by the clustering method were significantly smoother and less erratic compared to the single template approach.
  • Low FP Rates: The clustering approach consistently kept FP rates below 10%, which is critical for avoiding unnecessary irradiation of healthy tissues.
  • Computational Efficiency: While not explicitly quantified in terms of time, the paper suggests clustering is computationally more efficient than the multiple template method due to reduced template numbers.

Limitations and Future Work

  • Dataset Size: The study used datasets from only five patients and a limited number of fractions.
  • Inter-fraction Variability: The impact of variations between treatment sessions (inter-fraction variability) was not extensively investigated, though it's expected to be minor as templates are derived from the same fraction.
  • Aperiodic Motion: While the method is designed to handle drifts, aperiodic intra-fraction motion was not extensively tested.
  • Future Directions:
    • Larger Datasets: Testing the algorithms on more extensive patient data and longer treatment durations.
  • Improved Reference Signals: Developing more robust methods for generating reference gating signals for validation.
  • Dosimetric Impact: Evaluating the dosimetric consequences of the current error levels to determine if further reduction is necessary.
  • Inhale Gating: Exploring the application of these methods for gating during the inhale phase, which could be beneficial for sparing organs like the heart.
  • Clinical Implementation: Moving towards clinical implementation of the developed methods.

Summary updated

Abstract

For gated lung cancer radiotherapy, it is difficult to generate accurate gating signals due to the large uncertainties when using external surrogates and the risk of pneumothorax when using implanted fiducial markers. We have previously investigated and demonstrated the feasibility of generating gating signals using the correlation scores between the reference template image and the fluoroscopic images acquired during the treatment. In this paper, we present an in-depth study, aiming at the improvement of robustness of the algorithm and its validation using multiple sets of patient data. Three different template generating and matching methods have been developed and evaluated: (1) single template method, (2) multiple template method, and (3) template clustering method. Using the fluoroscopic data acquired during patient setup before each fraction of treatment, reference templates are built that represent the tumour position and shape in the gating window, which is assumed to be at the end-of-exhale phase. For the single template method, all the setup images within the gating window are averaged to generate a composite template. For the multiple template method, each setup image in the gating window is considered as a reference template and used to generate an ensemble of correlation scores. All the scores are then combined to generate the gating signal. For the template clustering method, clustering (grouping of similar objects together) is performed to reduce the large number of reference templates into a few representative ones. Each of these methods has been evaluated against the reference gating signal as manually determined by a radiation oncologist. Five patient datasets were used for evaluation. In each case, gated treatments were simulated at both 35% and 50% duty cycles. False positive, negative and total error rates were computed. Experiments show that the single template method is sensitive to noise; the multiple template and clustering methods are more robust to noise due to the smoothing effect of aggregation of correlation scores; and the clustering method results in the best performance in terms of computational efficiency and accuracy.
BibTeX
@inproceedings{cui2007robust,
  author = {Ying Cui and Jennifer Dy and G Sharp and Brian M. Alexander and Steve Jiang},
  title = {{Robust fluoroscopic respiratory gating for lung cancer radiotherapy without implanted fiducial markers}},
  booktitle = {Physics},
  year = {2007},
  doi = {10.1088/0031-9155/52/3/015},
}
Figure