Return to contents page

II. Methods

A. Scanning Protocol

All scans were performed on a 21 slice Posicam PET scanner (Positron Corp., Houston, TX) equipped with a rotating rod transmission source. Our Rb-82 rest-stress cardiac protocol consisted of a resting emission scan, followed by a transmission scan, followed by a dipyridamole stress emission scan. The acquisition and processing steps were (briefly):

B. Screening

Based on an initial estimate for a processed dataset of 1 MB and a desire to limit disk usage to about 250 MB, we retrieved 252 patient studies performed between 10/93 and 5/94. For each patient, we retrieved the patient header (containing all of the demographics and processing information), the FBP reconstruction of the attenuation coefficients (AC), and the circumferential profile data (the pixel data and the radial position data). We excluded patients whose bodies extended beyond the FOV ( 75 patients, 30%):

and whose arms were in the FOV ( 10 patients, 4%):

leaving us with 167 patients.

C. Processing

The processing of a patient dataset consisted of 4 steps:

  1. Decimation

    The reconstructed AC images were decimated and scaled from 1.7x1.7x5.125 mm, 16-bit voxels to 6.8x6.8x5.125 mm, 8-bit integer voxels with units of 1000*cm-1 (an AC of 0.1 cm-1 has a pixel value of 100). This was done to reduce disk space requirements.

  2. Edge-Finding

    The body and lung edges were found in radial coordinates using a simple thresholding technique.

    This figure illustrates the steps involved:

    The following algorithm was applied on a slice-by-slice basis:

    1. Find the center of mass of the slice.

    2. Transform to polar coordinates about the center of mass using 64 angles over 360 degrees with 0.5 pixel (3.4 mm) bilinear-interpolated radial sampling.

    3. Find the outer edge of the body by thresholding and searching from the outside into the center on the polar map. The first pixel greater than 0.03 cm-1 is the outer edge of the body.

    4. Calculate the geometric center of the body outline and re-transform to polar coordinates about the geometric center.

    5. Find the outer edge of the body again.

    6. Find the inner edge of the lung by searching from the center to the edge of the body with a threshold of 0.06 cm-1. The first pixel less than this value is the inner edge of the lung.

    7. The radial position of the outer edge of the lung is obtained by counting the number of pixels between the inner lung and the outer edge of the body that are less than the lung threshold value (0.06 cm-1) and adding this number of pixels to the inner edge of the lung.

    8. The body, inner lung, and outer lung edges are saved for the slice as three arrays of 64 floating-point radii over 360 degrees.

  3. Edge-Averaging

    Since the radial distances were relative to the geometric centers of the bodies and all patients' body edges extended the full 360 degrees, the body edges of 2 or more patients could be averaged together trivially.

    The averaging of the inner and outer lung edges was a little tricky. We elected to use a 2-pass algorithm. In the first pass we found the average angularpositions of the right and the left lungs (start angle and stop angle). This required that we identify the edges as belonging to the right or the left lung. If there was no clear break between the lungs, they were separated at 90 degrees and 270 degrees. In the second pass we interpolated each patient's lungs to the average angular positions before averaging them with the other patients.

  4. Warping the Attenuation Coefficient Images

    Software was developed to warp (in 2 dimensions) a patient's AC images to a set of template edges. The template could be the average of a group of patients, another patient's edges, or any other set of edges.

    Eight regions (see figure below) were segmented from the patient's polar AC images and bilinearly warped into the corresponding regions of the template.

    The figure below shows a single slice from a patient's AC image, the template edges used (these particular edges are from an average of all 167 patients), and the resulting warped AC slice.

    Notice how the tissue in the middle of the right lung is put into region 3 of the template.

    By warping a set of patient AC images to a common template and averaging the warped AC images we can build models from groups of patients.

_______________________________

Results

Return to contents page

Last updated 6/12/97
(Had to use XBM images when I first did this, now GIF's can be transparent...)