Salomon Brülisauer

Bachelor Thesis
Supervisors: Dieter Schweizer, Prof. Orçun Göksel

Deep Learning based Rigid Registration of 3D Ultrasound Volumes

Rigid registration of ultrasound volumes can help to analyze volumes which were generated from a different position, angle of view and different point in time. It computes the affine transformation of the position and angle of the ultrasound transducer such that one can transform/sample one image and lay one upon the other. As base there is a neural network which accepts two volume patches and outputs the affine transformation. We use it as a subsystem when we build a system which is enabled to match whole volumes instead of patches. The algorithms, which coordinate the process in such a system, are the main part in this thesis. We will have a look at different approaches and discuss their pros and cons. First there is a probabilistic approach, then we study a graph algorithm, a learning based algorithm will be discussed next and finally an iterative method. What all algorithms use the same is a loss function which indicates the accuracy of specific transformation parameters. A detailed description for the setup of the working environment is given at the beginning. That is an attempt to reduce overhead for further investigations. This thesis shows how to extend and generalize the ability of registering fixed size patches to arbitrary size volumes (with a lower bound of the size).