Fluoroscopic tumor tracking for image-guided lung cancer radiotherapy

Tong Lin, Laura Cerviño, Xiaoli Tang, Nuno Vasconcelos, Steve JiangPhysics 2009 · DOI

TL;DR

This paper introduces a novel markerless fluoroscopic tumor tracking algorithm for lung cancer radiotherapy. By correlating the tumor's motion with the movement of surrounding anatomical features (surrogates) using Principal Component Analysis (PCA) and regression, the method achieves accurate real-time tracking without invasive markers, paving the way for more precise image-guided treatments.

What Problem Does This Paper Address?

Image-guided radiotherapy for lung cancer is crucial for delivering precise radiation doses while minimizing damage to healthy tissues. However, a major challenge is the constant motion of lung tumors due to respiration. Existing methods to manage this motion fall into a few categories:

  1. External surrogates: Using external markers (like skin markers or breathing belts) to infer internal tumor position. This approach is often inaccurate because the correlation between external and internal motion can vary daily or even during a single treatment session.
  2. Implanted fiducial markers: Placing small markers directly in or near the tumor. While more accurate, this method carries risks like pneumothorax (collapsed lung) and may not be widely accepted by patients.
  3. Markerless tracking: Directly tracking the tumor in fluoroscopic images without any implanted markers. Previous markerless methods often rely on template matching, which struggles when tumor boundaries are unclear or image quality is poor – a common issue in lung cancer imaging.

This paper aims to bridge this gap by developing a markerless, non-invasive, and accurate fluoroscopic tumor tracking method that overcomes the limitations of existing approaches.

Methods

The proposed algorithm employs a four-step approach:

  1. Surrogate ROI Selection: Regions of Interest (ROIs) are identified in the initial fluoroscopic image. These ROIs are placed on anatomical features (like the diaphragm or lung boundary) that are expected to move in correlation with the tumor. The tumor itself can also be an ROI if its shape is discernible.
  2. Spatiotemporal Pattern Extraction: Instead of directly tracking the ROIs, the algorithm extracts their motion patterns. This is achieved by using Principal Component Analysis (PCA) to reduce the dimensionality of the image data within each ROI. Each ROI's intensity map is projected into a low-dimensional space, creating a parametric representation (a set of coordinates) that captures its dynamic behavior. The paper uses the first three principal components for each surrogate ROI, resulting in a 9-dimensional feature vector.
  3. Regression Model Establishment: A regression model is built to capture the correlation between the tumor's position (ground truth) and the parametric representations of the surrogate ROIs. This step uses training data from the initial part of the fluoroscopic sequence.
  4. Tumor Position Prediction: In real-time, the established regression model uses the current parametric representations of the surrogates to predict the tumor's location.

The authors tested four regression methods: linear regression, two-degree polynomial regression, Artificial Neural Network (ANN), and Support Vector Machine (SVM).

Datasets

The study utilized retrospective fluoroscopic image sequences from ten lung cancer patients. For each patient, 15 seconds of images (225 frames) were used for training the algorithm, and the remaining data from the sequence were used for testing. The ground truth tumor positions were manually identified by expert observers.

Key Insights and Ideas

  • Leveraging Surrogate Motion: The core idea is that while the tumor itself might be hard to track, its motion is often predictable by observing the movement of more discernible anatomical structures nearby. This is a clever way to infer internal motion from external or semi-external cues without direct invasive markers.
  • PCA for Feature Representation: Using PCA to create low-dimensional, parametric representations of ROIs is key. This method effectively captures the essence of the motion within each ROI, making it robust to minor variations and reducing computational complexity compared to tracking raw pixel data.
  • Regression for Correlation: The regression step is crucial for translating the surrogate motion patterns into accurate tumor position predictions. The comparison of different regression techniques highlights the effectiveness of machine learning in this complex task.
  • Markerless Advantage: The algorithm's markerless nature is a significant advantage, potentially increasing patient acceptance and reducing procedural risks associated with fiducial marker implantation.

Key Results

The proposed algorithm demonstrated promising results:

  • Mean Tracking Error: An average localization error of 2.1 pixels.
  • Maximum Tracking Error (95% confidence): A maximum error of 4.6 pixels.
  • Pixel Size: The pixel size is approximately 0.5 mm, meaning the average error is around 1 mm, and the 95% confidence error is around 2.3 mm.
  • Robustness: ANN regression showed slightly better performance and was more robust, with a maximum 95% confidence error of 6.5 pixels, compared to other methods where this value could double.
  • Speed: The algorithm achieved a processing time of 8.2 ms per frame on a standard PC, which is fast enough for real-time tracking during radiotherapy.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Surrogate Selection: Currently, surrogate ROIs are selected manually. Future work aims to develop automatic and optimal surrogate selection methods, as the choice of surrogates significantly impacts prediction accuracy.
  • Number of Surrogates/Features: The optimal number of surrogates and PCA coefficients per ROI is not definitively determined. Further research is needed to find the best trade-off between prediction accuracy and computational complexity.
  • Regression Function Enhancement: The current regression relies on instantaneous surrogate positions. Future work could explore regression functions that incorporate the temporal smoothness of PCA trajectories for improved prediction.
  • Clinical Implementation: Performing the regression before treatment is ideal, but changes in tumor size or position might necessitate recalculating the regression model.
  • Image Views: The study primarily used anterior-posterior (AP) views due to the difficulty in identifying tumors in other views. Future work plans to use fiducial markers as ground truth for testing in different views.
  • 3D Tracking: The current method only extracts 2D tumor motion. To achieve full 3D tracking, integration with other imaging modalities (like 4D-CT) or development of methods to infer the third dimension from planar images is required.

Summary updated

Abstract

Accurate lung tumor tracking in real time is a keystone to image-guided radiotherapy of lung cancers. Existing lung tumor tracking approaches can be roughly grouped into three categories: (1) deriving tumor position from external surrogates; (2) tracking implanted fiducial markers fluoroscopically or electromagnetically; (3) fluoroscopically tracking lung tumor without implanted fiducial markers. The first approach suffers from insufficient accuracy, while the second may not be widely accepted due to the risk of pneumothorax. Previous studies in fluoroscopic markerless tracking are mainly based on template matching methods, which may fail when the tumor boundary is unclear in fluoroscopic images. In this paper we propose a novel markerless tumor tracking algorithm, which employs the correlation between the tumor position and surrogate anatomic features in the image. The positions of the surrogate features are not directly tracked; instead, we use principal component analysis of regions of interest containing them to obtain parametric representations of their motion patterns. Then, the tumor position can be predicted from the parametric representations of surrogates through regression. Four regression methods were tested in this study: linear and two-degree polynomial regression, artificial neural network (ANN) and support vector machine (SVM). The experimental results based on fluoroscopic sequences of ten lung cancer patients demonstrate a mean tracking error of 2.1 pixels and a maximum error at a 95% confidence level of 4.6 pixels (pixel size is about 0.5 mm) for the proposed tracking algorithm.
BibTeX
@inproceedings{lin2009fluoroscopic,
  author = {Tong Lin and Laura Cerviño and Xiaoli Tang and Nuno Vasconcelos and Steve Jiang},
  title = {{Fluoroscopic tumor tracking for image-guided lung cancer radiotherapy}},
  booktitle = {Physics},
  year = {2009},
  doi = {10.1088/0031-9155/54/4/011},
}
Figure