CAD provides value in spotting difficult 'flat' colonic lesions

Computer-aided detection may have a role in helping radiologists spot hard-to-find flat lesions on CT colonography, according to a pair of studies presented in March at the European Congress of Radiology.

One retrospective study, based on a Japanese cohort, found that a Medicsight CAD system had a sensitivity of 84.6% when tested on 26 known nonpolypoid early-stage cancers. Another retrospective study, based on 208 cases, found that a prototype Siemens CAD system had a 93%-plus sensitivity for flat lesions.

The effectiveness of CT colonography and optical colonoscopy as screening tools was called into question with the publication of a study in the Journal of the American Medical Association concluding that flat or depressed colorectal neoplasms are more likely than polyps to contain in situ or submucosal carcinoma.

"Nonpolyoid colorectal neoplasms are more difficult to detect because the subtle findings can be difficult to distinguish from those of normal mucosa," wrote Dr. Roy Soetikno and colleagues in JAMA (2008;299:1027-1035). The prevalence of nonpolypoid lesions in the JAMA study was 9.35% overall and 6.01% in the group undergoing colon cancer screening. In the total group of 764 patients with nonpolypoid lesions, the prevalence of cancer was 0.82%. It was 0.32% in the screening arm.

The studies at the ECR suggest that CAD shows promise in spotting these lesions, which may be missed in normal reading situations. As many as one-fifth of colonic lesions may have a flat morphology, said Dr. Stuart Taylor of the University College of London, lead author in the study that examined flat lesions from the Japanese cohort.

The rationale for the study was that CAD screening of polypoid tumors is increasingly robust, and reading errors in CT colonography are usually perceptual-radiologists just don't see the lesion-so CAD may have value in the identification of flat lesions, Taylor said.

The study subjected 30 known early T1 colon cancers to the CAD program at varying levels of sensitivity. Most of the cancers were considered to be IIa under the Paris classification system; morphologically they were at least twice as wide as they were high.

The CAD program's sensitivity filter used three settings ranging from perfectly spherical to relatively flat. The researchers ruled out six of the 30 lesions because they were spherical. Of the remaining 24, the lesion sizes ranged from 7 to 60 mm with a mean size of 25 mm.

At the lowest sphericity setting, the CAD system spotted 20 cancers (83.3%); at the medium setting, it spotted 17 cancers (70.8%); and at the highest setting, it spotted 13 (54.1%). False positives ranged from nine to 36 based on the filter setting.

Taylor attributed part of the program's success to the fact that the CAD system was able to find small protuberances on even the flat lesions.

As with any CAD program, there is a trade-off between the number of false positives and the likelihood of finding cancer, Taylor said.

The study of the prototype from Siemens tested 208 cathartic data sets from multiple institutions and databases. Cancers ranged in size from 6 to 25 mm, and all data sets were collected using multislice CT with four or more detectors.

CAD identified 71 polyps, more than half of them sessile or pedunculated, said Marcos Salganicoff, Ph.D., of Siemens Medical Solutions U.S.A. The researchers identified two other sets of flat polyps: flat, with a length to height ratio of greater than two, and shallow flat, with a height >3 mm. There were no depressed or perfectly flat lesions. The polyps were further identified as medium (6 to >10 mm) or large (=10 mm).

Among the sessile and pendunculated polyps, the CAD was 100% sensitive across both size categories (42). Among the 29 flat polyps, the sensitivity was 93.8% for medium-sized polyps (16) and 92.3% for large polyps (13). Among the 15 shallow flat polyps, the system was 92.9% sensitive for medium polyps (14) and 100% sensitive for large polyps (1). Median false positives were two per volume.