TU‐D‐204C‐02: Machine Learning in Real‐time Tumor Localization

R. li, S. Jiang

Research output: Contribution to journalArticle

Abstract

Real‐time tumor localization is an essential step to achieve conformai lung cancer radiotherapy. In practice, tumors may be localized either directly (e.g., in fluoroscopy) or indirectly by some type of surrogate. Direct localization with implanted markers poses the risk of pneumothorax to lung cancer patients, while indirect localization is limited by its accuracy. Direct localization in fluoroscopy without implanted markers achieves a reasonable tradeoff between accuracy and risk to the patients. However, due to the poor soft tissue contrast and superposition of different anatomical structures, it is generally very difficult to localize the tumor in fluoroscopy with conventional methods in computer vision and image processing. Recently, machine learning has gained great popularity in many aspects of radiation therapy. In this lecture, we will demonstrate the applications of various machine learning techniques in the context of real‐time tumor localization in lung cancer radiotherapy. These cover a wide range of well established machine learning techniques, including principal component analysis (PCA), artificial neural networks, and support vector regression, etc. We will distinguish between two different paradigms when applying these techniques: one in a two‐dimensional (2D) framework and the other in a three‐dimensional (3D) framework. In the 2D framework, a (supervised) regression model is built between a parametric representation of the fluoroscopic images. and the tumor locations of the fluoroscopicimag fluoroscopic images. Because the representation of the fluoroscopic images is implicit in the 2D framework, it is difficult to account for deformational and large translational changes in the tumor. In the 3D framework, by incorporating the prior information in 4DCT or 4DCBCT, a PCA model is constructed to explicitly represent the entire lung motion in a realistic and efficient way. This model allows one to use a single x‐ray projection image to not only derive the 3D tumor location, but also reconstruct the corresponding volumetric image of the patient. With the aid of graphics processing units, this computationally intensive task can be achieved within half a second given one projection. This lecture provides an overview of the different machine learning techniques used in real‐time tumor localization. A detailed description of these machine learning techniques is also presented. Finally, extensions to the current framework as well as some future directions are discussed. Learning Objectives: 1. Understand the benefits and challenges of real‐time tumor localization without implanted markers. 2. Understand the principles of different types of machine learning techniques for real‐time tumor localization. 3. Understand how machine learning techniques are used to achieve accurate real‐time tumor localization. Conflict of Interest: This work is partially supported by Varian Master Research Agreement.

Original languageEnglish (US)
Pages (from-to)3395-3396
Number of pages2
JournalMedical Physics
Volume37
Issue number6
DOIs
StatePublished - 2010

Fingerprint

Neoplasms
Fluoroscopy
Lung Neoplasms
Radiotherapy
Principal Component Analysis
Machine Learning
Conflict of Interest
Pneumothorax
X-Rays
Learning
Lung
Research

ASJC Scopus subject areas

  • Biophysics
  • Radiology Nuclear Medicine and imaging

Cite this

TU‐D‐204C‐02 : Machine Learning in Real‐time Tumor Localization. / li, R.; Jiang, S.

In: Medical Physics, Vol. 37, No. 6, 2010, p. 3395-3396.

Research output: Contribution to journalArticle

@article{40888dd3e52849938b05e8e94c317b6c,
title = "TU‐D‐204C‐02: Machine Learning in Real‐time Tumor Localization",
abstract = "Real‐time tumor localization is an essential step to achieve conformai lung cancer radiotherapy. In practice, tumors may be localized either directly (e.g., in fluoroscopy) or indirectly by some type of surrogate. Direct localization with implanted markers poses the risk of pneumothorax to lung cancer patients, while indirect localization is limited by its accuracy. Direct localization in fluoroscopy without implanted markers achieves a reasonable tradeoff between accuracy and risk to the patients. However, due to the poor soft tissue contrast and superposition of different anatomical structures, it is generally very difficult to localize the tumor in fluoroscopy with conventional methods in computer vision and image processing. Recently, machine learning has gained great popularity in many aspects of radiation therapy. In this lecture, we will demonstrate the applications of various machine learning techniques in the context of real‐time tumor localization in lung cancer radiotherapy. These cover a wide range of well established machine learning techniques, including principal component analysis (PCA), artificial neural networks, and support vector regression, etc. We will distinguish between two different paradigms when applying these techniques: one in a two‐dimensional (2D) framework and the other in a three‐dimensional (3D) framework. In the 2D framework, a (supervised) regression model is built between a parametric representation of the fluoroscopic images. and the tumor locations of the fluoroscopicimag fluoroscopic images. Because the representation of the fluoroscopic images is implicit in the 2D framework, it is difficult to account for deformational and large translational changes in the tumor. In the 3D framework, by incorporating the prior information in 4DCT or 4DCBCT, a PCA model is constructed to explicitly represent the entire lung motion in a realistic and efficient way. This model allows one to use a single x‐ray projection image to not only derive the 3D tumor location, but also reconstruct the corresponding volumetric image of the patient. With the aid of graphics processing units, this computationally intensive task can be achieved within half a second given one projection. This lecture provides an overview of the different machine learning techniques used in real‐time tumor localization. A detailed description of these machine learning techniques is also presented. Finally, extensions to the current framework as well as some future directions are discussed. Learning Objectives: 1. Understand the benefits and challenges of real‐time tumor localization without implanted markers. 2. Understand the principles of different types of machine learning techniques for real‐time tumor localization. 3. Understand how machine learning techniques are used to achieve accurate real‐time tumor localization. Conflict of Interest: This work is partially supported by Varian Master Research Agreement.",
author = "R. li and S. Jiang",
year = "2010",
doi = "10.1118/1.3469269",
language = "English (US)",
volume = "37",
pages = "3395--3396",
journal = "Medical Physics",
issn = "0094-2405",
publisher = "AAPM - American Association of Physicists in Medicine",
number = "6",

}

TY - JOUR

T1 - TU‐D‐204C‐02

T2 - Machine Learning in Real‐time Tumor Localization

AU - li, R.

AU - Jiang, S.

PY - 2010

Y1 - 2010

N2 - Real‐time tumor localization is an essential step to achieve conformai lung cancer radiotherapy. In practice, tumors may be localized either directly (e.g., in fluoroscopy) or indirectly by some type of surrogate. Direct localization with implanted markers poses the risk of pneumothorax to lung cancer patients, while indirect localization is limited by its accuracy. Direct localization in fluoroscopy without implanted markers achieves a reasonable tradeoff between accuracy and risk to the patients. However, due to the poor soft tissue contrast and superposition of different anatomical structures, it is generally very difficult to localize the tumor in fluoroscopy with conventional methods in computer vision and image processing. Recently, machine learning has gained great popularity in many aspects of radiation therapy. In this lecture, we will demonstrate the applications of various machine learning techniques in the context of real‐time tumor localization in lung cancer radiotherapy. These cover a wide range of well established machine learning techniques, including principal component analysis (PCA), artificial neural networks, and support vector regression, etc. We will distinguish between two different paradigms when applying these techniques: one in a two‐dimensional (2D) framework and the other in a three‐dimensional (3D) framework. In the 2D framework, a (supervised) regression model is built between a parametric representation of the fluoroscopic images. and the tumor locations of the fluoroscopicimag fluoroscopic images. Because the representation of the fluoroscopic images is implicit in the 2D framework, it is difficult to account for deformational and large translational changes in the tumor. In the 3D framework, by incorporating the prior information in 4DCT or 4DCBCT, a PCA model is constructed to explicitly represent the entire lung motion in a realistic and efficient way. This model allows one to use a single x‐ray projection image to not only derive the 3D tumor location, but also reconstruct the corresponding volumetric image of the patient. With the aid of graphics processing units, this computationally intensive task can be achieved within half a second given one projection. This lecture provides an overview of the different machine learning techniques used in real‐time tumor localization. A detailed description of these machine learning techniques is also presented. Finally, extensions to the current framework as well as some future directions are discussed. Learning Objectives: 1. Understand the benefits and challenges of real‐time tumor localization without implanted markers. 2. Understand the principles of different types of machine learning techniques for real‐time tumor localization. 3. Understand how machine learning techniques are used to achieve accurate real‐time tumor localization. Conflict of Interest: This work is partially supported by Varian Master Research Agreement.

AB - Real‐time tumor localization is an essential step to achieve conformai lung cancer radiotherapy. In practice, tumors may be localized either directly (e.g., in fluoroscopy) or indirectly by some type of surrogate. Direct localization with implanted markers poses the risk of pneumothorax to lung cancer patients, while indirect localization is limited by its accuracy. Direct localization in fluoroscopy without implanted markers achieves a reasonable tradeoff between accuracy and risk to the patients. However, due to the poor soft tissue contrast and superposition of different anatomical structures, it is generally very difficult to localize the tumor in fluoroscopy with conventional methods in computer vision and image processing. Recently, machine learning has gained great popularity in many aspects of radiation therapy. In this lecture, we will demonstrate the applications of various machine learning techniques in the context of real‐time tumor localization in lung cancer radiotherapy. These cover a wide range of well established machine learning techniques, including principal component analysis (PCA), artificial neural networks, and support vector regression, etc. We will distinguish between two different paradigms when applying these techniques: one in a two‐dimensional (2D) framework and the other in a three‐dimensional (3D) framework. In the 2D framework, a (supervised) regression model is built between a parametric representation of the fluoroscopic images. and the tumor locations of the fluoroscopicimag fluoroscopic images. Because the representation of the fluoroscopic images is implicit in the 2D framework, it is difficult to account for deformational and large translational changes in the tumor. In the 3D framework, by incorporating the prior information in 4DCT or 4DCBCT, a PCA model is constructed to explicitly represent the entire lung motion in a realistic and efficient way. This model allows one to use a single x‐ray projection image to not only derive the 3D tumor location, but also reconstruct the corresponding volumetric image of the patient. With the aid of graphics processing units, this computationally intensive task can be achieved within half a second given one projection. This lecture provides an overview of the different machine learning techniques used in real‐time tumor localization. A detailed description of these machine learning techniques is also presented. Finally, extensions to the current framework as well as some future directions are discussed. Learning Objectives: 1. Understand the benefits and challenges of real‐time tumor localization without implanted markers. 2. Understand the principles of different types of machine learning techniques for real‐time tumor localization. 3. Understand how machine learning techniques are used to achieve accurate real‐time tumor localization. Conflict of Interest: This work is partially supported by Varian Master Research Agreement.

UR - http://www.scopus.com/inward/record.url?scp=85024818365&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85024818365&partnerID=8YFLogxK

U2 - 10.1118/1.3469269

DO - 10.1118/1.3469269

M3 - Article

AN - SCOPUS:85024818365

VL - 37

SP - 3395

EP - 3396

JO - Medical Physics

JF - Medical Physics

SN - 0094-2405

IS - 6

ER -