Over-segmentation of 3D medical image volumes based on monogenic cues

Abstract

In this thesis a novel approach to compute a 3D over-segmentation (supervoxels) for radiological image data is presented. It allows to cope with the high levels of noise and low contrast encountered in clinical data such as Computed Tomography (CT), Optical Coherence Tomography (OCT) and Magnetic Resonance (MR) images. The method, MonoSLIC, employs the transformation of the image content to its monogenic signal as primal representation of the image. The phase of the monogenic signal is invariant to contrast and brightness and by selecting a kernel size matched to the estimated average size of the superpixels it highlights the locally most dominant image edge. Employing an agglomeration step similar to the one used in SLIC-superpixels yields superpixels/-voxels with high fidelity to local edge information while being of regular size and shape. The proposed approach is compared to state of the art over-segmentation methods on the real-world images of the 2D Berkley Segmentation Dataset (BSD) converted to gray-scale, as well as on challenging 3D CT and MR volumes of the VISCERAL dataset. It yields a highly regular, robust, homogeneous and edge-preserving over-segmentation of the image/volume while being the fastest approach. For 3D volumes the method is 3 times faster than the state of the art. Due to its invariance to contrast and brightness it yields 11% higher recall rate when dealing with MR and CT volumes. There is also no parameter that needs to be tuned, increasing the usability of the method.