Registration-based segmentation of thoracic 4D MRI and respiratory motion model for lung cancer radiotherapy.

• Main Author: He, Shuai.
• Other Authors: Poh Chueh Loo
• Format: Final Year Project
• Language: English
• Published: 2013
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/54671
• Institution: Nanyang Technological University

Lung cancer is one of the world’s top killers, responsible for many deaths every year. A common treatment for lung cancer is radiotherapy, which usually requires careful planning prior to the treatment to improve the accuracy of radiotherapy and thus reduce side effects. Four-dimensional magnetic resonance imaging (4D-MRI) technique is a promising tool for radiotherapy planning due to its ability to generate high soft-tissue contrast images and capture variation of respiratory motion over multiple breathing cycles without the use of radiation. Automatic 4D registration-based image segmentation can be performed to analyze 4D MR images, which requires a set of semi-automatically segmented images to serve as the initial template for the automatic scheme and the ground truth for validation. The objectives of this study are: firstly, to improve the level of control over the semi-automatic image segmentation process by optimizing active contour evolution parameters; secondly, to improve the registration performance by determining the optimum step size; and thirdly, to model the relative diaphragm position against time for motion prediction as well as to study the effect of tumour on lung motion. Through this study, it was observed that snake evolution based on intensity regions behaved better than that based on image edges. As the two parameters of intensity-regions-based snake tool, namely balloon force and curvature force increase, the stopping iteration right before overshooting of the active contour raises from a minimum of 35 to a maximum of 185, which indicates an increase in level of control during the process. From this study, it was also learned that when the step size of the registration scheme is equal to or more than 2, the performance of the system becomes extremely unstable. The optimum step size was found out to be 1.5, where the dice similarity coefficient (DSC) is the highest, tolerance (TOL) is the lowest, and the stabilizing iteration is small. It was also found out that while simple 1st order sinusoidal model can be utilized to model controlled-breathing relative diaphragm position with R-square values higher than 0.85, 4th order sinusoidal model is optimum for modeling free-breathing data in this study. The effect of tumour on lung motion could be considered insignificant in terms of alteration in respiratory phase and period. The amplitude of lung motion does not show a direct relationship with the size of tumour. This could lead to the hypothesis that the range of motion for lungs is affected by a combination of various factors, including the stage of cancer.