J. Michael Tyszka

Department of Radiology, Cedars-Sinai Medical Center, 8700 Beverly Blvd., Los Angeles CA 90048

Received 12/16/96; Accepted 12/24/96; On-line 01/01/96

2. METHODS

All CT data was acquired using a HighSpeed Spiral CT scanner (General Electric Medical Systems, Milwaukee, WI) equipped with a power injector and automated contrast injection timing software. Vascular images were acquired during the arterial phase of contrast administration. Bronchial images were acquired during one or two breath-holds. Slice thickness, pitch and field of view were optimized based on radiological, hardware and dosimetric requirements. Typical slice thicknesses were in the range 1.5mm to 3mm with a slice pitch between 1.5mm and 3mm. All images were originally 512 x512 pixels in size with a field of view in the range 28cm to 36cm.

All post-processing was performed on a SPARCstation 20 MP with a ZX graphics accelerator (Sun Microsystems, Inc. Mountain View, CA.). CT images were transferred to the workstation by means of a 4mm Digital Audio Tape (DAT) and imported into the AVS processing environment (Advanced Visual Systems, Inc. Waltham, MA.) which provides many image processing operations and allows custom operations to be written and implemented quickly. Preliminary processing involved reduction of the slice resolution of the CT images from 512 x 512 to 256 x 256 by voxel averaging and cropping the data set to the volume of interest.

A point within the structure to be segmented (such as the bronchial airway) was selected manually. A region was then "flooded" from this starting point with boundaries defined by a manually selected maximum and minimum image intensity. The resulting set of voxels defined all connected elements in the data set with an image intensity in the predefined range. This algorithm was implemented as a custom operation within the AVS environment.

The region was then passed to the isosurface module provided by AVS and a mesh surface defined for viewing and checking the segmented structure. The region was alaso passed to a custom written module which implemented the "marching-cubes" algorithm (60 and generated triangular mesh surface information in VRML form as a text file. The text file (with a .wrl suffix) could then be linked to a web page and disseminated across the Internet for remote viewing.

In this implementation, the true spatial proportions of the data were preserved only approximately by interpolation of the data points. Consequently, quantitative measuements of distance, area or volume from the VRML datasets are not possible. Future versions of the VRML module will include this feature.

Figure 1. Simple image segementation was achieved by "flooding" a region from a seed point using an image intensity range as a constraint. The resulting set of points described a connected volume corresponding to a given material (in this case air within the tracheobronchial tree). No adjustments were made for partial volume effects.
Figure 2. The results of applying the marching cubes algorithm to the segmented data of the previous figure. Marching cubes generates a triangular mesh surface representing the boundary of the segmented volume. The Gouraud shading algorithm generates a smooth, lit version of the mesh surface. The mesh information is easily converted to VRML form for Internet dissemination.