Dual-energy subtraction improves CAD accuracy for lung nodule detection

Overlying bone structure in the chest can stump computer-aided detection systems on the lookout for pulmonary nodules. Digitally subtracting the bone images from the chest radiograph can significantly improve CAD performance, according toMaryland researchers.

One obstacle to both radiologists and CAD systems in lung nodule detection has been the superimposition of anatomic structures on the radiograph, said lead author Dr. Eliot L. Siegel, chief of radiology and nuclear medicine at the University of Maryland/Baltimore VA Medical Center.

"Conventional chest radiographs have had a disappointingly low sensitivity in detecting pulmonary nodules," he said. "CAD has been shown to improve that sensitivity."

Dual-energy subtraction has also been shown to improve both sensitivity and specificity in lung nodule detection. Siegel and colleagues examined the possible synergies between dual-energy subtraction and CAD.

They used a CAD program that had been trained on conventional radiographs without dual-energy subtraction acquisition. Using a combination of conventional chest radiography, dual-energy subtraction chest radiography, and thoracic CT, they examined 26 patients with previously detected benign or suspicious lung nodules.

The study employed radiologist consensus as well as comparison CT images of the thorax as a gold standard. Four nodules between 10 and 15 mm in diameter were found.

In the conventional radiographs, the CAD program did not detect any of the four nodules. The system also suggested 59 marks on the images. In comparison, the program detected three out of the four nodules in the soft-tissue images that had had the overlying bone structure digitally subtracted. The system made only 39 marks on these images.

"Even though the sample size was small, this was a statistically significant difference," Siegel said.

Because the investigators used a CAD system that had not been trained on dual-energy subtracted radiographs, accuracy could be improved even further by tailoring CAD training, according to Siegel.

The researchers studied only dual-energy subtraction and CAD technologies from single vendors, Siegel said. Future studies should investigate the interplay between dual-energy subtraction acquisition parameters and CAD using CT correlation and examine the possible differences between algorithms from different vendors.