Multi-modal matching of 2D images with 3D medical data

Image registration is the process of aligning images of the same object taken at different time points or with different imaging modalities with the aim to compare them in one coordinate system. Image registration is particularly important in biomedical imaging, where a multitude of imaging modalities exist. For example, images can be obtained with X-ray computed tomography (CT) which is based on the object’s X-ray beam attenuation whereas magnetic resonance imaging (MRI) underlines its local proton density. The gold standard in pathology for tissue analysis is histology. Histology, however, provides only 2D information in the selected sections of the 3D tissue. To evaluate the tissue’s 3D structure, volume imaging techniques, such as CT or MRI, are preferable. The combination of functional information from histology with 3D morphological data from CT is essential for tissue analysis. Furthermore, histology can validate anatomical features identified in CT data. Therefore, the registration of these two modalities is indispensable to provide a more complete overview of the tissue. Previously proposed algorithms for the registration of histological slides into 3D volumes usually rely on manual interactions, which is time-consuming and prone to bias. The high complexity of this type of registration originates from the large number of degrees of freedom. The goal of my thesis was to develop an automatic method for histology to 3D volume registration to master these challenges.
The first stage of the developed algorithm uses a scale-invariant feature detector to find common matches between the histology slide and each tomography slice in a 3D dataset. A plane of the most likely position is then fitted into the feature point cloud using a robust model fitting algorithm.
The second stage builds upon the first one and introduces fine-tuning of the slice position using normalized Mutual Information (NMI). Additionally, using previously developed 2D-2D registration techniques we find the rotation and translation of the histological slide within the plane. Moreover, the framework takes into account any potential nonlinear deformations of the histological slides that might occur during tissue preparation.
The application of the algorithm to MRI data is investigated in our third work. The developed extension of the multi-modal feature detector showed promising results, however, the registration of a histological slide to the direct MRI volume remains a challenging task.