Towards Physiological-based Respiratory Motion Modeling: Improvement on 4D Image Quality and Feature-based Deformable Image Registration

Thumbnail Image



Journal Title

Journal ISSN

Volume Title

Repository Usage Stats



Purpose: To develop methods towards physiological-based respiratory motion management, through the enhancement of image quality of four dimensional magnetic resonance imaging (4D-MRI), as well as the deformable image registration (DIR) accuracy utilizing the data commonly available in radiotherapy treatment.

Methods: First part of the work is on the development of a probability density function (PDF) based method to improve the image quality of 4D-MRI. The overall idea is to identify a few main breathing cycles (and their corresponding weightings) that can best represent the main breathing patterns of the patient and then reconstruct a 4D image set for each of the identified main breathing cycles. This method is implemented in three steps: (1) The breathing signal is decomposed into individual breathing cycles, characterized by amplitude, A, and period, T; (2) Individual breathing cycles are grouped based on values of A and T to determine the main breathing cycles. If a group contains more than 10% of all breathing cycles in a breathing signal, it is determined as a main breathing pattern group and is represented by the average of individual breathing cycles in the group; (3) For each main breathing cycle, a set of 4D images is reconstructed using a result-driven sorting method adapted from our previous study. The probability-based sorting method was first tested on 26 patients’ breathing signals to evaluate its feasibility of improving tumor motion displacement PDF. Breathing motion PDFs were calculated for the original breathing signal (reference PDF), the average single-cycle breathing cycle (PDFsingle), and the main cycles determined by the probability-based sorting method (PDFprob). Similarity of PDFsingle/PDFprob to the reference PDF was evaluated using the Dice Similarity Coefficient (DSC) between the areas under the PDFsingle/PDFprob and under the reference PDF, noted as DSCsingle/DSCprob. The new method was subsequently tested for a sequential image acquisition scheme using the 4D digital eXtended Cardiac Torso (XCAT) phantom. Performance of the sorting methods was evaluated in terms of “tumor” volume precision and accuracy as measured by the 4D images, and also the accuracy of AIP of the 4D images.

The second part of the work is on the development of a radial basis function (RBF)-based hybrid DIR framework utilizing sparsely available measured motion information. The framework is carried out in three steps: (1) A base, intensity-based displacement vector field (DVF) is converted to an intensity-based coefficient matrix comprising expansion coefficients for the Wendland’s RBF; (2) The intensity-based coefficient matrix is modified under the guidance of sparely distributed measured motion information to generate the hybrid coefficient matrix; (3) The hybrid coefficient matrix is converted to the hybrid DVF. The hybrid DIR framework was tested on an in-house built lung motion phantom and compared with five different DIR methods including Velocity, MIM, ILK, OHS, and Elastix. Intensity-based and hybrid DVFs were compared to the known DVFs of the phantom. Synthesized EOI (primary) images were generated by deforming the phantom EOE (secondary) images with the intensity-based/hybrid DVFs. The intensity-based/hybrid synthesized EOI images were compared with the phantom EOI images. The effects of the number of landmarks were also studied by implementing the hybrid DIR process with 60-600 landmarks in the lungs.

The third part of the work is on the development of a method to automatically extract physiological landmarks (lung vessel bifurcations) using a novel random walking strategy. The lung images were first pre-processed with segmentation, binarizaion, and skeletonization. After the pre-processing, a vessel tree skeleton was acquired. For bifurcation extraction, first, end voxels were detected. Then a tracer would start from an end voxel and randomly move to one of its neighbors until reaching another end voxel. Numbers of being passed by the tracer were recorded for each voxel and a voxel was determined as a bifurcating voxel if it was visited more than all of its neighbors. The method was evaluated on a 7-layer simulated bifurcating tree, generated based on a modified Holton’s method for generation of botanical trees.

Results: For the first part, improved similarity of breathing motion PDF to reference PDF was observed as compared to PDFsingle, indicated by the significant increase in DSC (DSCprob = 0.89±0.03, and DSCsingle = 0.83±0.05, p-value <0.001). Based on the simulation study on XCAT, the probability-based method outperforms the conventional phase-based methods in qualitative evaluation on motion artifacts and quantitative evaluation on tumor volume precision and accuracy and accuracy of AIP of the 4D images.

In the second part, DIR errors were reduced in all tested cases after applying the hybrid DIR framework. Intensity-based only vs hybrid 3D DIR errors per voxel are 1.27 mm vs 1.07 mm, 1.56 mm vs 1.55 mm, 6.33 mm vs 4.81 mm, 3.53 mm vs 2.94 mm, and 2.00 mm vs 1.67 mm for for Velocity, MIM, ILK, OHS, and Elastix respectively. Reduction in intensity difference between the phantom EOI images and the synthesized EOI images are also observed in the hybrid synthesized EOI images. Study on the impacts of the number of landmarks confirms that a larger amount of landmarks can improve the accuracy of the hybrid DVF, and also indicates that the hybrid DIR framework may result in a larger improvement to intensity-based DVFs of poorer accuracy.

In the third part, an algorithm based on modified Holton’s model has been implemented to generate bifurcating vessel trees for testing the automatic bifurcation extraction method. Based on the results on a 7-layer bifurcating tree, the extraction method is able to detect more than 85% of the ground-truth bifurcations with 95% of the detected bifurcations located within 2 voxels from the nearest ground-truth bifurcation. The detection rate rises to 90% of the maximum detection rate in 4 iterations. The percentage of correctly detected bifurcations remains stable for different number of iterations ranging from 1 to 100. However, the variation of the percentage increases as the number of iterations lowers.

Conclusion: In this thesis, we developed a number of novel methods on the improvement in 4D image quality and hybrid deformable image registration, creating a more solid ground toward physiological plausible respiratory motion modeling. In particular, we have developed a probability-based multi-cycle sorting method for improved 4D image reconstruction and demonstrated the feasibility of a novel probability-based multi-cycle 4D image reconstruction method, presenting potential advantages over the conventional phase-based reconstruction method for radiation therapy motion management. In addition, we developed a hybrid DIR framework which is capable of potentially reducing registration errors by utilizing sparsely distributed measured motion information. The flexibility of the framework is expected to enhance the DIR accuracy in radiotherapy research. Furthermore, we have developed an automatic method to extract vessel bifurcations based on the random walking strategy and demonstrated its high accuracy and efficiency in a preliminary evaluation.






Liang, Xiao (2017). Towards Physiological-based Respiratory Motion Modeling: Improvement on 4D Image Quality and Feature-based Deformable Image Registration. Dissertation, Duke University. Retrieved from


Dukes student scholarship is made available to the public using a Creative Commons Attribution / Non-commercial / No derivative (CC-BY-NC-ND) license.