Computer-aided detection proves useful in lung x-rays

Radiologists can maximize their diagnoses of lung abnormalities using computer-aided detection systems if they develop a better understanding of the strengths and shortcomings of every factor involved in the process, according to studies presented at the RSNA meeting.

"Detecting early lung cancer remains a significant challenge on chest x-rays, because such cancers are often small, subtle, or obscured by overlapping structures. Missed lung cancers on chest x-rays are second only to breast cancer for malpractice actions," said Dr. Charles S. White, chief of thoracic radiology at the University of Maryland Medical Center in Baltimore.

White released a retrospective review performed on an imaging database of 2100 patients diagnosed with lung cancer between 1993 and 2006. The research team assessed the utility of CAD to detect lesions overlooked in these patients' final reports.

The investigators found that CAD could detect many lesions overlooked by human readers on chest radiography. Although size, location, and conspicuity did not generally affect the frequency of detection, CAD could miss lesions larger than the software is set to catch, he said.

The researchers found missed lung cancers in 88 patients. These lesions were mostly peripheral and ranged in size from 0.5 cm to 5.5 cm. CAD identified 33% and 37.7% of undetected lesions on a per-film and per-patient basis, respectively. CAD detected peripheral lesions just as frequently as central ones, providing an average rate of 4.5 false-positive results per radiograph.

In a separate study, Dr. David P. Naidich, a professor of radiology and medicine at New York University, presented results on a retrospective review of 200 chest CT studies from four different institutions. Naidich and colleagues assessed CAD's added value in identifying pulmonary nodules according to size and conspicuity. Seventeen general radiologists reviewed a first round of all blinded studies with additional CAD-supplied findings. Five chest radiologists performed a blinded final round of reads.

The investigators found that CAD could bolster lung nodule detection as a second reader depending on the nodules' size and conspicuity. The more conspicuous the lesion, the easier it was for CAD to identify. These conclusions became tricky, however, since all the radiologists involved in the reading process defined conspicuity in different ways, Naidich said.

A third study, conducted by Dr. Justus Roos, a radiologist at Stanford University, applied CAD to multislice CT scans from 20 patients with lung cancer. Roos and colleagues found that time played a major role in how radiologists performed with CAD. An initial period of upward performance in lung nodule detection became significantly degraded over time, as radiologists tended to become complacent over false positives, Roos said.

By adding CAD, the evaluation initially provided a sharp rise in sensitivity with a minimal increase in false-positive detections for a period of about 90 seconds. This period gave way to an abrupt flattening of reader performance and a 4% increase in false-positive detection due to increased acceptance of false-positive CAD hits. The diagnostic performance decline observed during this transition was statistically significant. The transition zone occurred after a mean of 15 CAD hits and after a mean CAD evaluation time of 90 seconds.

"The interaction between humans and CAD is actually a very important topic. A lot of research has to be done to optimize this interface," Roos said. "Understanding the complex interactions between CAD and radiologist interpretation may lead to practice guidelines that optimize the benefits of CAD during lung CT interpretation."