SAN FRANCISCO - When a radiologist sees a possible lung nodule on a chest x-ray, he or she begins to visualize the image without normal anatomic structures in order to see the suspicious area more clearly. That process, known as "visual subtraction," is the idea behind a new method of reducing false positives in computer-aided diagnosis (CAD) of lung nodules.
At the 14th annual Computer Assisted Radiology and Surgery conference in San Francisco on June 29, Hiro Yoshida, Ph.D., from the University of Chicago radiology department, said the new CAD scheme was able to reduce false positives by nearly half.
Although CT is considered the most sensitive modality for detecting lung nodules, chest x-rays prevail for initial lung screening due to their low cost and low radiation dose.
Yet radiologists miss up to 30% of lung nodules imaged on chest x-rays, so various CAD schemes have been developed to improve detection rates. Many of the schemes produce a high number of false-positive nodules, Yoshida said, so an effective method of removing them is needed in order to make the process practical.
"Our studies show that if you have a large number of false positives from the computer output, then radiologists aided by computer output tend to have worse results (than with unaided detection)," Yoshida said.
When radiologists see a suspicious region of interest (ROI) on a chest x-ray, they tend to look at the corresponding region on the opposite side of the image to determine normal anatomy, then visualize the region of interest without the normal structures, Yoshida said. The Chicago team translated this "visual analysis strategy" using "local contralateral subtraction" into a computer algorithm and a CAD scheme.
The contralateral subtraction process works as follows: For each ROI, the program identifies a corresponding contralateral ROI about four times as large, then examines each pixel in the search area to determine the correlation between the suspicious ROI and the new ROI, Yoshida said.
The ROI with the greatest correlation is then flipped in the left-right direction and used as the contralateral ROI. The contralateral ROI is then extracted from the suspicious ROI, effectively "removing the normal background structures, while keeping the abnormal structures," Yoshida said.
A nonlinear image mapping registration method is then applied to compensate for natural variations between the ROIs. This process distorts the contralateral ROI to match the suspicious ROI, which is then subtracted from the image to yield a residual ROI with the anatomic structures removed.
Following this process, a true-positive ROI will have a region of pixels at the center that has a higher average intensity than the surrounding region, Yoshida said. Signal-to-noise ratios, which tend to be lower for nodules than for false positives, are used to delineate and identify abnormal structures.
Yoshida's team evaluated a total of 200 radiographs that contained 550 ROIs, including 51 nodules and 499 false positives. With a signal-to-noise ratio threshold of 4.0, the method correctly identified 222 of the 499 false positives as negative, resulting in the elimination of 44% of the false-positive findings, without eliminating true positives.
"This strategy is potentially useful in improving the performance of CAD schemes for the detection of lung nodules," Yoshida said.
By Eric Barnes
AuntMinnie.com staff writer
July 4, 2000
Copyright © 2000 AuntMinnie.com
