Imaging gets a helping hand: image processing, chest and colon CAD
SCAR U tackled computer assisted interpretation of images in a three-part series on Thursday featuring Bradley Erickson, MD, PhD, of Mayo Clinic (Rochester, Minn.) who addressed practical image processing and visualization; Matthew Brown, PhD, of University of California Los Angeles, who discussed theories and practicalities of chest CAD; and Ronald Summers, MD, PhD, of the National Institutes of Health Clinical Center (Bethesda, Md.) who handled theories and practicalities of colon CAD.

The Image Processing and Visualization Primer

Image processing offers multiple tools to better extract and display information, enabling radiologists to slice, dice, enhance and measure images. It also facilitates new viewing modes. Tools fall into four categories: basic rendering, pre-processing and enhancement, morphologic operators and classification and measurement, Erickson said.
Basic rendering applications include surface rendering, which applies colors to identified objects, and volume rendering techniques such as maximum intensity projections (MIPs), gradients and sums. MIP is a familiar, frequently used tool that requires minimal pre-processing. MIP, however, does not provide quantitative data and can be misleading. Other challenges in rendering include noise, but to reduce noise, radiation dose or imaging time must be increased.
Pre-processing and enhancement tools include a variety of linear filters and newer non-linear filters. A median filter orders all pixels in a sample and applies the median value to reduce noise. An anisotropic diffusion filter models the diffusion of water to retain edges and reduce noise. Morphologic filters can be used to remove noise or unwanted objects. Morphologic erosion can remove structures, or dilation can be used to add layers to separate structures and improve 3D images.
Classification assigns labels to parts of an image, while segmentation breaks an image into the largest parts. Both tools can help obtain baseline and follow up measurements.
"Computer tools will become every day practice and provide the ability to make diagnoses not possible with visual review," Erickson predicted. On the downside, image processing tools are not standard among vendors and the correct tool and its use is not always clear.

Theories and Practicalities of Chest CAD

Brown led a whirlwind tour of the chest CAD arena, not only covering the basics but also focusing on system evaluation and integration. "Chest CAD is poised to play a pivotal role in the future of radiology," Brown opined.
Lung CAD is primarily used for lung module detection and is driven by increased interest in screening and early diagnosis as well as the data explosion generated by multislice CT scanners. The goal of CAD is to identify nodules with high sensitivity, providing the radiologist with a second opinion.
CAD relies on image segmentation to separate nodules from normal anatomy. It also extracts quantitative features and classifies nodules.
Brown listed a few challenges for chest CAD:
  • Small vessels can appear as nodules
  • Irregular nodules, ground glass nodules and other nodules present a challenge
  • All systems involve a trade-off between detection and false positives
Chest CAD is available in both x-ray and CT versions. The x-ray version can detect nodules as small as five millimeters on chest x-rays; however, it performs best with 10 to 30 millimeter nodules. CT systems detect nodules four millimeters or greater. Systems may include additional tools for quantitative assessment, attenuation statistics, automatic tracking and 3D visualization.
Sites evaluating chest CAD can analyze systems based on standalone performance or changes in reader performance. Brown reported, "Studies have demonstrated that reader sensitivity improves from 40 to 60 percent pre-CAD to 75 to 90 percent post-CAD."

On the downside, chest CAD does increase false positives, which can force radiologists to take an unnecessary second look at an area and possibly initiate additional studies for the patient. Moreover, significant inter-reader variability exists, with CAD boasting its largest performance gains among less experienced radiologists.
"We need further studies on patient outcomes to determine if significantly fewer patients die from lung cancer with CAD as a second reader," Brown said.
He also reported varying implementations with some vendors marketing complete CAD workstations and another selling CAD software. Moreover, DICOM CAD output is not readable on PACS workstations and not all CAD systems store results on PACS. "The next logical step is more complete integration into the PACS workflow," Brown said. Other future steps include new applications such as CAD for diffuse lung disease.

Theories and Practicalities of CAD for virtual colonoscopy

Clinically, colorectal cancer stands as the second leading cause of cancer mortality. Screening does decrease incidence, and the gold standard for screening is optical colonoscopy, which has a sensitivity for polyps and cancers larger than one centimeter in the 87 to 95 percent range.
Virtual colonoscopy or CT colonography relies on a CT scan of the colon. "The rationale for colon CAD is to avoid perceptual errors and reduce inter-observer variability," Summers explained.
Colonoscopy CAD is designed to identify features characteristic of colon cancer, and clinical trials report a sensitivity of 65 to 100 percent depending on the patient mix and polyp size, with one to seven false positives per patient.
Recent advances in colon CAD include polyp segmentation to measure the internal characteristics of polyps and distinguish true and false polyps. Ileocecal valve detection accounts for one-fourth of all false positives, and efforts to identify the ileocecal valve and rectal tube may enhance performance, says Summers. Other work focuses on registering supine and prone studies for more accurate viewing of identified polyps and CAD of polyps covered by opacified fluid.