This thesis presents segmentation-enhanced registration methods to support image-guided treatments of brain tumors. The solutions are proposed for magnetic resonance imaging (MRI) and ultrasound acquisitions.
Multiparametric MRI data are acquired before and after neurosurgery. To accurately detect any pathological remnants or regrowth, a comparison of successive acquisitions is often performed, which can be further improved by the application of automatic registration methods. Such solutions are required to address the lack of one-to-one correspondence between pre- and post-operative acquisitions, which is often caused by the presence of pathology, and the need to identify the optimal MRI sequence to guide the registration process. This thesis proposes two automatic solutions for registering pre- and post-operative volumes, an iterative method and a deep learning-based approach. In both algorithms, segmentation masks can be used to select the voxels of non-corresponding pathological tissues, whose contribution is discarded from the registration process. A quantitative analysis of the impact of pathology exclusion on the registration methods is conducted. Furthermore, an evaluation of the influence of multiple MRI sequences on the registration result is presented. In addition, a deep learning method for the segmentation of resection cavities in multi-parametric MRI volumes is proposed. The influence of different MRI sequences on the segmentation of these structures is evaluated. The masks generated by the segmentation method are used in the aforementioned registration solutions.
During tumor resection procedures, multiple intraoperative ultrasound volumes are acquired at different stages to provide continuous imaging of the ongoing surgery. However, due to the brain shift, which is a displacement of brain tissue from its preoperative position, there is a need to compare successive acquisitions to track the changes during the surgery. Automatic registration methods can be employed to facilitate this task, but they have to account for the lack of correspondence between successive acquisitions. Therefore, this thesis proposes two iterative segmentation-enhanced registration methods to compensate for the brain shift in intra-patient ultrasound volumes. The first method involves automatically segmenting masks of corresponding healthy structures, which are then used to guide the registration of successive acquisitions. In the second method, masks of non-corresponding resection cavities are segmented, and subsequent ultrasound volumes are registered by excluding the contribution of the segmented structures from the search for correspondences.