Preview

PDF, English Download (26MB) | Terms of use

Abstract

Multimodal image registration benefits the diagnosis, treatment planning and the performance of image-guided procedures in the liver, since it enables the fusion of complementary information provided by pre- and intrainterventional data about tumor localization and access. Although there exist various registration methods, approaches which are specifically optimized for the registration of multimodal abdominal scans are only scarcely available. The work presented in this thesis aims to tackle this problem by focusing on the development, optimization and evaluation of registration methods specifically for the registration of multimodal liver scans. The contributions to the research field of medical image registration include the development of a registration evaluation methodology that enables the comparison and optimization of linear and non-linear registration algorithms using a point-based accuracy measure. This methodology has been used to benchmark standard registration methods as well as novel approaches that were developed within the frame of this thesis. The results of the methodology showed that the employed similarity measure used during the registration has a major impact on the registration accuracy of the method. Due to this influence, two alternative similarity metrics bearing the potential to be used on multimodal image data are proposed and evaluated. The first metric relies on the use of gradient information in form of Histograms of Oriented Gradients (HOG) whereas the second metric employs a siamese neural network to learn a similarity measure directly on the image data. The evaluation showed, that both metrics could compete with state of the art similarity measures in terms of registration accuracy. The HOG-metric offers the advantage that it does not require ground truth data to learn a similarity estimation, but instead it is applicable to various data sets with the sole requirement of distinct gradients. However, the Siamese metric is characterized by a higher robustness for large rotations than the HOG-metric. To train such a network, registered ground truth data is required which may be critical for multimodal image data. Yet, the results show that it is possible to apply models trained on registered synthetic data on real patient data. The last part of this thesis focuses on methods to learn an entire registration process using neural networks, thereby offering the advantage to replace the traditional, time-consuming iterative registration procedure. Within the frame of this thesis, the so-called VoxelMorph network which was originally proposed for monomodal, non-linear registration learning is extended for affine and multimodal registration learning tasks. This extension includes the consideration of an image mask during metric evaluation as well as loss functions for multimodal data, such as the pretrained Siamese metric and a loss relying on the comparison of deformation fields. Based on the developed registration evaluation methodology, the performance of the original network as well as the extended variants are evaluated for monomodal and multimodal registration tasks using multiple data sets. With the extended network variants, it is possible to learn an entire multimodal registration process for the correction of large image displacements. As for the Siamese metric, the results imply a general transferability of models trained with synthetic data to registration tasks including real patient data. Due to the lack of multimodal ground truth data, this transfer represents an important step towards making Deep Learning based registration procedures clinically usable.