Segmentation of lungs from volumetric CT-scan images using prior knowledge (shape and texture)

Abstract

This thesis presents a hierarchical segmentation scheme for The segmentation of lungs from volumetric CT images that concerns variational segmentation methods, namely geodesic active surfaces (GAS) and active surfaces without edges (ASWE), a volumetric similarity registration technique, statistical shape modelling using principal component analysis (PCA), and volumetric texture modelling.

GAS and ASWE are 3-D extensions of their 2-D version, geodesic active contours (GAC) and active contours without edges (ACWE). The two models are generalized into a unified framework, referred to as integrated active contours (IAS). Numerical implementation methods are derived for 3-D and the experiments are conducted both in 2-D and 3-D on synthetic and CT images. Global and local properties of active contours/surfaces under different parameter settings are presented and several applications of these models are proposed based on experimental results.

The similarity registration technique aims tom find an optimal match between shapes with respect to rotation, scale and translation parameters. In this registration method, PCA is initially employed to calculate the principal axes of shapes. These principal axes are used to obtain a coarse match between shapes to be registered. Then geometric moments are exploited to estimate the isotropic scale parameter. The rotation and translation parameters are estimated by phase correlation techniques which take advantage of the fast Fourier transform (FFT). Experimental results demonstrate that the proposed technique, compared with the standard iterative gradient descent method, is fast, robust in the presence of severe noise, and suitable in registering various types of topologically complex volumetric shapes.

Shape decomposition using PCA is the current state of the art and is widely drawn on in building deformable shape templates. The major problem to be solved in the modelling is to find proper PCA shape parameters that best approximate a novel shape of the same class. A comparison of popular methods for parameter estimation in the literature is presented and a hybrid coarse-to-fine method based on previous works is proposed.

The method achieves satisfactory accuracy over previous works and is validated by a database of lung shapes.

A hierarchical shape-based segmentation method that incorporates GAS, ASWE, similarity registration, and statistical shape modelling is proposed to extract lungs from volumetric low-dose CT images. The method is extensively experimented with a large variety of images including synthetic images with noise and occlusions, low-dose CT images with artificial noise and synthetic tumors, and a low-dose CT database. The results indicate that the method is robust against noise and occlusions.

Last but not least, a novel volumetric texture modelling technique based on isotropic Gaussian Markov random field (IGMRF) is developed and applied to low-dose CT images of lungs. Based on the proposed texture modelling, a hard classification approach is suggested to provide proper initializations for the shape-based segmentation method and enables the segmentation to achieve a higher degree of automation. The method is evaluated by low-dose CT images with synthetic tumors and the low-dose CT database. The experimental results suggest its suitability for offering proper initializations for shape-based segmentation.

More information

Identifiers

Catalogue record

Export record