Computer-aided diagnosis in contrast-enhanced CT colonography: an approach based on contrast.

Abstract

Automation in colorectal mass detection is achieved as soon as the voxels representing colorectal masses can be automatically segmented. We tested the Hounsfield (HU) value in intensely contrast enhanced high-resolution CT colonography for automated segmentation of colorectal masses in 18 patients with 35 polyps (28 < 10 mm, 7 > or = 10 mm) and 7 carcinomas. Mean HU values of the colonic wall and masses were determined to deduce a gradient threshold for a segmentation process, which encodes the voxels bordering the colonic lumen with a colour ranging in intensity from 0 to 100% according to the selected gradient threshold range in the volume rendering. The results of the automated segmentation process were superimposed on a virtual double contrast and endoscopic display and validated through correlation with morphology. Mean HU values and their standard deviations for the colonic wall, polyps < 10 mm, polyps > or = 10 mm and carcinomas were 63 +/- 24, 154 +/- 38, 116 +/- 41 and 108+/-29 HU, respectively. A gradient threshold ranging from 90 to 160 HU resulted in colour pools in 6 of 7 of polyps > or = 10 mm, and 6 of 7 carcinomas that were significant enough to prospectively guide the reader's attention to these masses. Due to the superposition of "false-positive" voxels in the projection view, the virtual endoscopic perspective was superior to the virtual double contrast display for controlling the segmentation results. The HU value is promising for automated segmentation of colorectal masses but needs to be combined with morphological parameters to render automated colorectal mass detection more accurate. Further refinements of the method with subsequent analysis of its accuracy, as well as analogue studies with contrast-enhanced MRI, appear warranted. More information at http://www.multiorganscreening.org