To evaluate the performance of automated computer-aided detection (CAD) of polyps in minimal-preparation (non-cathartic) CT colonography (mpCTC).
METHOD AND MATERIALS
35 patients underwent CTC (2.5 mm slice thickness, 1.25 mm reconstruction interval) in supine and prone positions (70 datasets) with a variety of iodine-based minimal preparations. Starting from the CT images, the CAD scheme extracted a thick 3D region encompassing the colon, including any regions enhanced by oral contrast, by use of a centerline-based colon extraction algorithm. Polyp candidates were detected from the extracted region(s) by use of rotation-invariant 3D shape features. In addition to air-filled regions, the CAD scheme also searched for polyp candidates submerged within regions of contrast-enhanced residual fluid with sufficient image quality, as indicated by a local evaluation by 3D texture features. False-positive polyp detections were reduced by use of a multidimensional shape and texture analysis based on a Bayesian neural network. The CAD scheme was trained on 121 patients who underwent CTC with standard pre-colonoscopy cleansing, as well as on a colon phantom which underwent CTC without and with a variety of stool-tagging preparations.
The 35 patients included 5 patients with 5 colonoscopy-confirmed polyps >=5 mm, including a 9-mm polyp submerged in residual fluid. When tested with the 35 patients, the CAD scheme yielded 100% by-patient and by-polyp detection sensitivities with 1.8 and 2.4 false positives per dataset on average, respectively. Most (55%) of the false-positive detections were caused by untagged or poorly tagged stool.
The CAD performance on mpCTC compares favorably with that of cathartically cleansed CTC. CAD may be useful in improving the diagnostic performance of mpCTC which is expected to improve patient compliance and comfort over cathartic preparations.
M.E.Z.: Recipient of a research grant from Amersham, Inc.