Accurate and Robust Centerline Extraction from Tubular Structures in Medical Images

Abstract

Extraction of centerlines is useful in interactive navigation and analysis of objects in medical images, such as the lung, bronchia, blood vessels, and colon. Given the noise and other imaging artifacts that are present in medical images, it is crucial to use robust algorithms that are (1) accurate, (2) noise tolerant, (3) computationally efficient, and (4) preferably do not require an accurate segmentation. We propose a new centerline extraction method that employs a Gaussian type probability model to estimate the boundaries of medical objects. The model is computed using an integration of the image gradient field. Probabilities assigned to boundary voxels are then used to compute a more robust distance field, that is less sensitive to noise. Distance field algorithms are then applied to extract the centerline. Noise tolerance of our method is demonstrated by adding Gaussian, Poisson and Rician noise to these datasets, and comparing results to traditional distance field based methods. Accuracy of our method was measured using two datasets with known centerlines, (1) a synthetically generated sinusoidally varying cylindrical dataset, and (2) a radiologist supervised segmented head MRT angiography dataset. Average errors for the cylinder dataset using our method was 0.5-0.8 voxels vs. 0.7-2.0 voxels using the traditional distance transform method; for the MRT dataset, it was 0.5-0.7 voxels vs 2.0-3.0 voxels for the traditional method. Additionally, experiments with six datasets were performed, (1) a second head MRT angiography dataset, (2) an aneurysm dataset, and (3) four colon datasets. Results of our approach illustrate the robustness of our centerline extraction method, in terms of the smoothness as well as reduced artifacts, such as spurious branches. Finally, the stability of our centerline is evaluated by measuring its sensitiveness to initialization and segmentation parameters (on the head MRT dataset), and found to vary on the average between 0.2-0.4 voxels. Running times of our algorithm are on the order of 1-7 minutes for datasets ranging from 256 × 256 × 256 to 409 × 409 × 219 voxels.