Template-driven segmentation of confocal microscopy images

Abstract

High quality 3D visualization of anatomic structures is necessary for many applications. The anatomic structures first need to be segmented. A variety of segmentation algorithms have been developed for this purpose. For confocal microscopy images, the noise introduced during the specimen preparation process, such as the procedure of penetration or staining, may cause images to be of low contrast in some regions. This property will make segmentation difficult. Also, the segmented structures may have rugged surfaces in 3D visualization. In this paper, we present a hybrid method that is suitable for segmentation of confocal microscopy images. A rough segmentation result is obtained from the atlas-based segmentation via affine registration. The boundaries of the segmentation result are close to the object boundaries, and are regarded as the initial contours of the active contour models. After convergence of the snake algorithm, the resulting contours in regions of low contrast are locally refined by parametric bicubic surfaces to alleviate the problem of incorrect convergence. The proposed method increases the accuracy of the snake algorithm because of better initial contours. Besides, it can provide smoother segmented results in 3D visualization.

Introduction

High quality 3D visualization of anatomic structures can facilitate diagnosis, surgical planning, biomedical inspection, structural understanding, etc. The raw data acquired from conventional imaging modalities such as computed tomography (CT), magnetic resonance imaging (MRI) and confocal microscopy are sequential 2D images. A volume dataset can be obtained by stacking up these images. Many rendering software packages provide 3D visualization of the volume dataset, but they cannot segment the different anatomic structures. An automatic segmentation process is necessary for practical applications.

Since a volume dataset can be obtained by stacking up sequential parallel 2D images, segmentation of the confocal microscopy images can be performed in 2D or 3D manner. For 3D segmentation, atlas-based and simplex-mesh-based approaches are two of the most popular methods. Atlas-based segmentation methods [1], [2], [3], [4], [5] find the transformation relationship between the atlas and the target data set by linear or nonlinear registration, and then warp a pre-segmented atlas to the target data set to perform segmentation or labeling. Nonlinear registration usually provides better segmentation result than linear registration, but it is very time-consuming and more sensitive to noise. For confocal microscopy images, noise is introduced during the specimen preparation process, such as the procedure of penetration or staining. Some regions of the images are of low contrast and have weak edges, which makes nonlinear local registration less reliable. Simplex-mesh-based methods [6], [7], [8] perform segmentation by capturing the features in the volume data with a closed spline surface. This kind of methods can provide smooth segmentation results in 3D, but cannot preserve sharp-features. In order to overcome this problem, a complex manual procedure for parameter assignment is necessary. When we have to preserve sharp-features, like in the Drosophila mushroom bodies, simplex-mesh-based methods are not favorable.

For 2D segmentation, the methods based on active contour models are the most popular. They are also known as the snake models [9], [10], [11], [12], [13], [14]. Energy-minimizing spline curves are found to capture the image features. When the edges are not connected, these approaches can still provide smooth and closed contours without the requirement of edge-linking. The initial contours affect the converged results of the active contour models, and they should be close to the object boundary to ensure correct convergence. The approach also encounters obstacles while dealing with confocal microscopy images. Some contour segments cannot capture the correct edges in regions of low contrast. Although the converged contours are still smooth curves, the reconstructed surface model of the segmentation result may have rugged surfaces.

This paper presents a specific segmentation algorithm for confocal microscopy images. It segments the confocal scans in 2.5D manner. The atlas-based segmentation with linear registration is adopted to provide a rough segmentation result. After model slicing, the resulting contours are regarded as the initial contours of the active contour models. The converged active contours are locally refined by parametric bicubic surfaces when contour segments are converged in regions of low contrast. The process of contour refinement can alleviate the problem of incorrect convergence. The proposed approach can increase the accuracy of the snake algorithm owing to better initial contours. It can also provide a better segmentation result for smoother surface reconstruction.