VC CAD system finds polyps in opacified fluid

Researchers from the National Institutes of Health in Bethesda, MD, have tweaked their colon CAD algorithm to detect polyps submerged, or partly submerged, in opacified fluid. The study, published in the January 2005 issue of American Journal of Roentgenology, represents another key step in the rapid evolution of colon CAD, paving the way for automated polyp detection with the use of oral contrast agents, which may have improved sensitivity in recent VC studies.

"With the administration of small amounts of oral contrast agents, residual fluid and feces become identifiable, wrote Dr. Ronald Summers and colleagues. "Bowel opacification, however, introduces a new challenge for computer-aided detection (CAD) of polyps on CT colonography because most CAD is designed to find polyps only in the air-filled colon" (AJR, January 2005, Vol. 184:1, pp. 105-108).

The CAD system consists of a colon segmentation algorithm, a fluid subtraction algorithm, and a CAD scheme, the authors explained. The study cohort consisted of 17 asymptomatic patients ages 40-79, taken from a much larger study group. All 17 patients had a colonoscopically proven polyp submerged in contrast-enhanced fluid. In all there were 22 submerged polyps among the 17 patients, including 14 measuring 0.5-0.9 cm, seven measuring 1-2 cm, and one measuring 5.1 cm, the authors wrote.


The patients followed a clear liquid diet, and underwent a cathartic bowel cleansing regimen consisting of 500 mL of dilute CT barium 2.1% solution (Scan C, Lafayette Pharmaceuticals, Yorba Linda, CA), 120 mL of diatrizoate meglumine and diatrizoate sodium solution (Gastroview, Mallinckrodt, Hazelwood, MO; or Gastrografin, Bracco Diagnostics, Princeton, NJ), 90 mL of sodium phosphate (Fleet 1 preparation, Fleet Pharmaceuticals, Lynchburg, VA) in divided doses, and two bisacodyl tablets (Dulcolax, Boehringer Ingelheim, Ingelheim, Germany).

The patients underwent scanning on a four- or eight-slice scanner (LightSpeed Plus or LightSpeed Ultra, GE Healthcare, Waukesha, WI) following patient-controlled colonic insufflation with room air, using 4 x 2.5-mm or 8 x 1.25-mm detector configuration, respectively, and 2.5-mm or 1.25-mm collimation, respectively, at 100 kVp and 100 mAs. CAD analysis was limited to the assessment of one dataset, either prone or supine, from each patient.

"We developed a colon segmentation method using a region-growing algorithm that, like a porpoise, "jumps" from air to fluid and back again until all portions of the colon are identified," Summers and colleagues wrote. The CAD sequence, they explained, consists of "region-growing segmentation with capability to traverse smoothly from air to opacified fluid and vice-versa; second, labeling of air and fluid regions and calculation of mean and SD of total colonic fluid CT attenuation; third, identification and labeling of air-fluid boundaries; and, fourth, second segmentation of fluid-filled segments to correct for possible leakage from the first segmentation."

The system distinguishes the findings in the data based on predefined thresholds for air (-800 H or lower) and fluid (276 H or higher), defining everything in between as colon wall.

And to prevent the relatively low CT value from causing "leakage" of the segmentation process into adjacent structures such as the small bowel, the researchers limited the thickness of the air-fluid boundaries to 2 voxels, also applying other geometric conditions.

"Then a ... modified fluid threshold is defined ... and the fluid regions are segmented once again with this new higher threshold," the authors wrote. The completed segmentation process defines every voxel in the dataset as air, fluid, air-fluid boundary, colonic wall, or other, and a modified isosurface procedure is used to draw the colonic wall.

Next the system begins looking for polypoid shapes by assessing the curvature of every vertex based on a convolution method and gradient calculations. It further applies a clustering process that winnows the findings down to those that pass certain predefined curvature tests; the survivors become candidate polyps. Each case was then assessed manually to ensure that segmentation was complete, the authors explained.

"To get a properly defined curvature for polyps located both in air and under fluid, we have to reverse the sign for the principal components of curvature calculated for vertices under fluid," they wrote.

The resulting endoluminal images clearly depicted haustral folds and a smooth colonic wall after fluid subtraction, according to the authors. Segmentation was complete in all cases. A mean of one seed was required to segment the colon (range 1-4, 1.6 ± 1.0 SD), and additional seeds were required only when the colon was collapsed, resulting in disconnected colon segments.

Six of 22 polyps were located on air-fluid boundaries, and 19 (86%) of the 22 polyps were detected by CAD, the authors wrote. There were only three false-positives, including two polyps measuring 5 mm and a hyperplastic polyp on an air-fluid boundary measuring 7 cm.

"We used a mixture of oral contrast medium that included both barium sulfate and an ionic iodine solution," Summers and colleagues wrote. "Barium sulfate is used for its ability to coat the lumen surface and tag fecal remnants. The ionic iodine solution is a water-soluble agent that mixes with the residual colonic fluid and increases the contrast between the fluid and the adjacent bowel wall. It is used to distinguish between fluid and soft tissue for colon surface rendering."

The goal is to separate the colonic wall from everything else in the dataset, including water, residual fecal material, and small-bowel data, to facilitate the detection of colonic wall lesions.

The resulting CAD system was able to detect polyps in both air- and water-filled portions of the bowel, as well as polyps that straddled both, the group concluded.

"Our CAD algorithm could be enhanced by setting an initial threshold to classify most of the fluid and then use local analysis to refine the classification," they wrote. "Future work will address optimization of the algorithm and reporting overall sensitivity and false-positive rates for larger numbers of patients."

By Eric Barnes staff writer
January 3, 2005

Related Reading

Low-prep VC study finds CAD can be fooled, December 23, 2004

Mass appeal: VC CAD doesn't stop at polyps, October 10, 2003

Knowledge-guided segmentation improves polyp detection, October 18, 2002

Virtual colonoscopy: 2-D vs. 3-D primary read, June 3, 2002

Prepless virtual colonoscopy shows early promise, June 3, 2002

Copyright © 2005

Page 1 of 655
Next Page