aalto1 untyped-item.component.html
Deep learning-based affine and deformable 3D medical image registration
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Major/Subject
Mcode
SCI3095
Degree programme
Language
en
Pages
82+8
Series
Abstract
In medical image registration, medical scans are transformed to align their image content. Traditionally, image registration is performed manually by clinicians or using optimization-based algorithms, but in the past few years, deep learning has been successfully applied to the problem. In this work, deep learning image registration (DLIR) methods were compared on the task of aligning inter- and intra-patient male pelvic full field-of-view 3D Computed Tomography (CT) scans. The multistage registration pipeline used consisted of a cascade of an affine (global) registration and a deformable (local) registration.
For the affine registration step, a 3D ResNet model was used. The two deformable methods that were investigated are VoxelMorph, the most commonly used DLIR framework, and LapIRN, a recent multi-resolution DLIR method. The two registration steps were trained separately; For the affine registration step, both supervised and unsupervised learning methods were employed. For the deformable step, unsupervised learning and weakly supervised learning using masks of regions of interest (ROIs) were used. The training was done on synthetically augmented CT scans. The results were compared to results obtained with two top-performing iterative image registration frameworks. The evaluation was based on ROI similarity of the registered scans, as well as diffeomorphic properties and runtime of the registration.
Overall, the DLIR methods were not able to outperform the baseline iterative methods. The affine step followed by deformable registration with LaPIRN managed to perform similarly to or slightly worse than the baseline methods, managing to outperform them on 7 out of 12 ROIs on the intra-patient scans. The inter-patient registration task turned out to be challenging, with none of the methods performing well consistently. For both tasks, the DLIR methods achieve a very significant time speedup compared to the baseline methods.