Duke researchers combine DR tomosynthesis with 3D CAD

Jun 3, 2007

The use of radiography to detect lung disease has been eclipsed in recent years by CT, which offers improved spatial resolution and the ability to find smaller tumors at an earlier stage. But a group from Duke University in Durham, NC, hopes to make radiography competitive again with the use of digital x-ray tomosynthesis and a computer-aided detection (CAD) algorithm.

CAD is already being used in CT and digital radiography, but as in mammography CAD, the number of false positives that are generated is a concern. False positives were a major reason behind the negative findings produced in a study of CAD's impact on mammography screening that was published earlier this year in the New England Journal of Medicine (April 5, 2007, Vol. 356:14, pp. 1399-1409).

The Duke researchers believe they can make CAD more accurate and reduce false positives by taking multiple low-dose digital radiography projection images that are analyzed by the CAD algorithm, then combined into a 3D view and analyzed again. They discussed their technique, called correlation imaging, in the May issue of the American Journal of Roentgenology.

The limits of chest radiography

Past studies on the use of radiography for lung cancer screening have found that the modality frequently misses subtle lung lesions, the authors said. This can be attributed to the "inherent imaging characteristics" of chest radiography, in which the "standard posteroanterior radiograph reduces the 3D anatomy of the chest to a 2D shadow." Thus, some anatomical features can obscure others, such as solitary pulmonary nodules (AJR, May 2007, Vol. 188:5, pp. 1239-1245).

"When a projection radiograph of the chest is taken from a different angle, however, it is less likely that the same anatomic features would overlap so as to obscure the same true nodules, or to produce false nodules in the same location as in the posteroanterior radiograph," the authors said.

The authors developed a prototype digital radiography system with a flat-panel detector that they said was similar to the XQ/i DR system by GE Healthcare of Chalfont St. Giles, U.K. The system produced images in a 2048 x 2048 matrix in 14 bits with 0.2-mm pixel pitch.

Tomosynthesis imaging requires multiple projections taken at different angles, so the researchers modified the system with a motorized gantry that enabled the x-ray tube to move along a vertical axis during image acquisition. Projection images were acquired in 11 seconds spaced evenly over a 20° angle range. The system's beam setting was set to 120 kVp and 5 mAs, and the combined correlation images produced radiation exposures ranging from 16% to 100% of the total exposure of a single posteroanterior chest radiograph.

Home-grown CAD

For the CAD portion of the study, the group used a CAD algorithm developed at Duke. The algorithm first analyzed each 2D projection image using a number of different criteria, such as difference-of-gaussians (DOG) filters. The CAD results from multiple projections were then combined to produce a 3D correlation image in which CAD results that correlated across projections reinforced each other. While the algorithm operates in 3D, results are displayed as markings on projection posteroanterior images, similar to other CAD algorithms.

To test the correlation imaging approach, the researchers applied the algorithm to both phantoms and a small group of seven human subjects who had lung cancer confirmed on CT. The human patients had a total of 18 nodules 1-21 mm in size, with an average size of 8.1 mm and a median size of 7 mm. The group collected 71 projection images in the phantom studies and 61 images for the human group.

In analyzing their results, the group found correlation imaging to be "a robust technique leading to a notable reduction in false-positive findings." Correlation imaging reduced false positives in the phantom cases by as much as 79% when the algorithm was set to produce a sensitivity of 65%, and 78% in the human cases.

The researchers found they were able to reduce the number of false positives by increasing the number of digital projections used in the 3D images. With the CAD algorithm set at an 88% sensitivity threshold, the algorithm produced 40 false positives with two projection images in the group of phantom images, but that dropped to 14 false positives when 17 projections were used. With the system set to 65% sensitivity, it produced 11 false positives with two projections and just three false positives with 17 projections.

This phenomenon tended to produce decreasing returns at about seven projections, leading the authors to conclude that five to seven projections could be the optimal number. A lower number of projections would result in lower radiation dose delivered to patients, the authors said.

Phantoms versus humans

While the system performed well in limiting false positives for the phantom studies, it produced "many more false-positive findings" in the human cases -- at a 65% sensitivity level and collecting seven projection images, it produced three false positives in the phantom versus 50 for humans. The higher number in humans occurred because some of the false-positive findings correlated with normal nodule-like anatomic features, the authors said.

The authors compared their prototype system to the only radiographic CAD system on the U.S. market, the RapidScreen RS-2000 developed by Riverain Medical of Miamisburg, OH (formerly Deus Technologies). The authors said that with the RapidScreen system performing at a 66% sensitivity level, it produced 5.3 false positives per image, according to the application the company filed to get the system approved by the U.S. Food and Drug Administration. That compares to the Duke system's three false-positive findings per case with 17 phantom projections and 25 false-positive findings with 15 human subject projections at the 65% sensitivity threshold.

The authors said that their results were limited by a number of factors, such as the small number of human cases, the use of a relatively simple CAD algorithm without feature extraction, the CAD reconstruction algorithm, and other issues. Despite the drawbacks, however, they were optimistic about the future of correlation imaging.

"The current correlation algorithm yields improved performance over single-view CAD," the authors concluded. "This finding indicates the potential of this technique for greatly enhancing identification of solitary pulmonary nodules in chest radiography."

By Brian Casey
AuntMinnie.com staff writer
June 4, 2007

CAD performs well in lung nodule detection, February 5, 2007

CAD turns in mixed performance for pulmonary embolism, January 16, 2007

Pilot study: Entropy-driven CAD zips through vast breast image database, August 2, 2006