Location: MIC-Arena, Maria-Goeppert-Str. 3, 23562 Lübeck
Title: Contrast- and modality-invariant self-similarity descriptors for deformable image registration
Abstract: The robustness and accuracy of image registration methods depends heavily on calculating similarity between images. In medical imaging, a linear relationship between intensities across scans rarely exists, ruling out simple metrics, which depend directly on intensity differences. Intensity distortions between scans can be caused e.g. by non-functional relationships of intensities across modalities, local change in image contrast, image noise or artefacts. Previous approaches have focussed on explicitly modelling the functional or statistical relationship of intensities using mutual information or cross-correlation. Furthermore, the use of scalar structural image representations based on gradient orientation or entropy has been explored.
In this seminar, I will present a novel approach that is based on local image descriptors, which are related to key-point descriptors used for interest points matching. In contrast to previous work, our self-similarity descriptors are evaluated densely for each voxel and are invariant to multi-modal images, not only to monotonic grayscale transformations. We achieve a high computational efficiency by using descriptor quantisation and Hamming distances. The effectiveness of this approach will be demonstrated for a number of challenging medical image registration tasks, including intra-operative ultrasound to MRI registration and motion estimation for inhale-exhale lung CT scans.
