Malin Larsson presents her Master's Thesis on Thursday 23/5 at 13.15 in MH:228

Abstract:

One way to reduce the growing workload for radiologists is to automate the transformation of MR images of the brain into the AC-PC coordinate system. In this thesis, both non-deformable and deformable image registration methods were used to register images to a template in the desired space, the AC-PC coordinate system. First, three non-deformable transformations were used together with three different templates and two similarity metrics. To transform an image to the AC-PC coordinate system, one only needs the rotation matrix and translation vector. Thus, the transformation matrix and the landmarks (AC, PC and MS) were used to extract the rotation matrix and translation vector from non-deformable and deformable image registration, respectively. The combinations of transformation, template and similarity metric were evaluated with four different metrics: distance to AC, angle in sagittal, coronal, and axial plane. The non-deformable transformations were rigid, similarity and affine transformation, while the templates were Colin27 T1- and T2-weight, and MNI305. The two similarity metrics used were mattes mutual information and cross correlation. The best combination in the non-deformable case was affine transform with Colin27 with T1-weight and mattes mutual information. The average angular error for this combination was 3.45°, 0.87° and 1.18° for the angles in the sagittal, coronal, and axial plane, respectively. The average distance error to AC was 3.11mm. The second part consists of the development of a deformable transformation that was semi-based on the first part, because it used the best template and the best similarity metric from the first part. The deformable transformation was symmetric normalization (SyN) with Colin27 with T1-weight and mattes mutual information. The average angular error for the deformable method was 1.84°, 1.24° and 1.12° for the angles in the sagittal, coronal, and axial plane, respectively. The average distance error to AC was 1.47mm. The average time for a non-deformable registration was 18 seconds, and 320 second for a deformable registration. Except for the time, the deformable method was on par with state-of-the-art machine learning methods from the literature. The non-deformable method performed worse than the methods from the literature, except when comparing the time for one registration.

Supervisors:  Anders Heyden, Grayson Webb (Sectra)

Examiner: Niels Christian Overgaard