Segmentation and editing of 3-dimensional medical images.

Abstract

Neuroradiologists rely on scanned images of the human brain to diagnose many pathologies. The images, even those collected in 3-dimensions, are typically displayed as a 2-dimensional collage of slices and much of the intrinsic 3-D structure of the data is lost. Image Atlases are commonly used to delineate and label Volumes Of Interest (VOIs) in 3-dimensional, slice-type, medical data sets. They can serve many purposes: to highlight important regions, to quantify the size and shape of structures in the images, to define a surface for 3-D rendering and to help in navigation through a series of images. To perform these functions, an individual atlas is required for each data set. The purpose of this thesis is to develop a link between the volume data and the individual atlas associated with each set of images. An automatic method of building an individual atlas from the volume data is proposed. The method uses a data-driven, bottom-up segmentation to produce a primitive atlas followed by a knowledge-driven, top-down merging and labelling stage to refine the primitive atlas into an individual atlas. The system was implemented in software using an object-oriented approach which allowed for a high quality user interface and a flexible and efficient implementation of the concepts of an atlas and a VOI. Tests were performed to judge the quality of the segmentations and of the atlas labellings. The results prove that the individual atlases created using the proposed method are sufficiently accurate to aid in visualizing 3-D structures in medical data sets and to quantify the sizes of these structures.