Método de contornos ativos radiais 3D: uma nova abordagem para segmentação de volumes

Abstract

Active Contour Methods (ACMs) are image segmentation techniques that consist of segmenting regions from a curve around the object that tends to mold to the edges of the object by minimizing an energy that is a function of the geometry of the curve and the intensity of the pixels of the image. There are many 2D and 3D ACMs. However, three dimensions segmentation tend to demand a high computational cost and among these techniques, the Radial Active Contour Method is the one that has the lowest computational cost but is limited to 2D space. In this thesis, a new three-dimensional Radial Active Contours Method (3DRACM) is proposed, which expands the concept of 2D radial ACMs and that analyzes information along beams (1D) that diverge on a plane from the center of the object. In three dimensions, the beams diverge in space with a combination of angulation (azimuth and elevation) from an internal point to the 3D object. Thus, despite three dimensions, the analysis continues to be along the beam (1D). These beams can be connected in different shapes and form different meshes or surfaces. These surfaces deform through energy equations to expand or contract until they reach the edges of the volume of interest. The main advantage of this new technique is its low computational complexity when compared to the literature techniques for 3D segmentation. To evaluate 3DRACM, the following metrics are used: position adjustment, size adjustment, shape adjustment and dice coefficient, in addition to calculation of computational cost. We performed tests on five types of synthetic volumes and five real chest TC scans. In these CT scans, we segmentat the lungs as a proof of concept of the new technique. The proposed technique is compared with the following techniques in the literature: 3D ACM, Morphological ACM and 3D Regions Growing. The results showed the efficiency of 3DRACM in the segmentation of the synthetic and real volumes with high correspondence rate in position, shape and size beyond expected low computational cost, producing results 16 times faster than the Morphological ACM and twice as fast as the 3D Region Growing in addition to having superior segmentation and no leaks.