PhD defence by José David Tascon Vidarte
This PhD defense will take place both physically and online. Click here to join online, via Zoom.
Real-time Deformable Image Registration Applied to Tumor Tracking in 2D Cine-MRI
Summary of thesis
Registration is the process of aligning a set of images and is considered a fundamental task in computer vision and medical image analysis. For medical applications where tissue is involved, registration requires a deformable transformation. The deformation implies more parameters to solve, and the optimization is computationally expensive. More efficient and high-performing deformable registration algorithms will benefit many applications such as image-guided radiotherapy and surgery.
This thesis aims to improve deformable image registration (DIR) performance and achieve real-time operation for tumor tracking in Cine-MRIs. We explore the state-of-the-art of DIR to select a strategy that accomplishes high performance. We choose to enhance intensity-based DIR algorithms with a variational (gradient-based) approach. The reasons are generalization capabilities, accuracy, and data availability. The research path to implement and validate our strategy is as follows. First, we explore how accurate intensity-based algorithms are with low contrast organs such as the liver. Next, we design a high-performance image registration library. Then, we obtained a fast algorithm with the lowest reported time in literature to solve DIR in real-time. Finally, we applied the DIR algorithm to tumor tracking in the context of image-guided radiotherapy.
A proper evaluation of a tumor tracking application requires considerable ground truth data. Therefore we design a novel Cine-MRI simulator that creates video sequences with the underlined delineation of the organ and the tumor. Furthermore, we use treatment Cine-MRIs with manual delineations. After the data preparation, we evaluate deformable image registration and other tracking algorithms with multiple organs. We find that DIR is more accurate in tracking organs compared to tumors. As a consequence of a multiple-organ strategy, we propose and validate how well the tracking algorithms replicate novel gating control signals for image-guided radiotherapy.
Finally, we made a comprehensive study on tumor tracking with nine algorithms to find the best solution. We include liver and lung patients from simulation and treatment data with the most challenging conditions. We propose a novel tracking method that combines template matching with deformable image registration. This algorithm was among the best-performing algorithms overall for tumor tracking. In summary, we find that the best algorithms perform close to interobserver variability and we prove that tracking tumors on lung and liver patients offers similar accuracy. All the code generated during the development of this thesis is publicly available.
- Associate Professor Bulat Ibragimov, Department of Computer Science, University of Copenhagen
- Professor Natasa Sladoje, Center for Image Analysis, Dept. of Information Technology, Uppsala University, Sweden.
- Associate Professor Line Katrine Harder Clemmensen, Technical University of Denmark, Compute
Associate Professor Sune Darkner, Image Analysis, Computational Modelling and Geometry Section, Dept. of Computer Science, UCPH.
Moderator for the defense
For a digital copy of the defense, go to https://di.ku.dk/english/research/phd/.