Radiology: Software boosts lung nodule detection
Software-processed (bone-suppressed) chest x-ray with nodule in right upper lung. Source: Radiology, April 2011.
The use of image post-processing software to suppress bones and locate lung nodules appears to significantly improve radiologists’ detection of lung cancer, while also producing additional false-positives on chest x-rays, according to a study published April 14 in Radiology.

A substantial majority of small and early stage lung cancers is associated with solitary pulmonary nodules. However, bone structures have been widely shown to obfuscate lung nodules on chest x-rays, with studies finding that bones commonly cause radiologists to miss cancers.

Several methods, including image post-processing software programs, have been employed to attempt to diminish the obstruction of bone in the identification of lung nodules. The authors tested a new software program that suppressed the ribs and clavicles, filtered noise and equalized contrast in an effort to improve radiologists’ detection of lung nodules and other lung cancer findings.

Fifteen radiologists, compensated by the software’s developer, read a total of 1,368 posterior chest x-rays without the software and with the software (in immediate sequence). The study, with the results under the control of the authors, was used by Riverain Medical Group to apply for FDA 510(k) product clearance.

Sixty-seven percent of images had no lung nodules, while the remaining 33 percent of images had a maximum of one nodule and no metastases, with all findings pre-confirmed by additional testing.

The mean area under the localized receiver operating characteristic (LROC) curve was 0.460 without the software and 0.558 when viewed with the program, a significant improvement in detection. This increase in area was relatively consistent for all 15 readers.

The software program likewise significantly enhanced reviewers’ sensitivity, from 50 percent to 66 percent. Meanwhile, the software significantly increased the average number of false-positives reported by radiologists, dropping specificity from 96 percent to 92 percent when using the post-processing software.

Moreover, the program led radiologists to change several correct cancer positions to incorrect locations or false-negative results.

“This detrimental effect of an increase of 2 percent in false-negative decisions with software should be compared with the much greater increase in the sensitivity for cancer detection of 16.8 percent,” offered Matthew Thomas Freedman, MD, MBA, and colleagues from the Lombardi Comprehensive Cancer Center and the department of oncology at Georgetown University Medical Center in Washington, DC.

Freedman and co-authors also observed that three-quarters of the lesions detected only while using the software were at least 70 percent obstructed by bone, implying that the software achieved its chief aim, to suppress bone and enhance visualization of covered lesions.

“This visualization software represents a new software approach to this problem and results in highly efficient removal of the bone component of chest images with few visible artifacts. It likely represents a clinically viable tool at this time,” the authors observed.

Freedman and colleagues noted that even with the software, the radiologists’ sensitivity was much lower than that found under lung CT scans. The authors also considered the possibility that, because the software was not used on an independent read, radiologists could have unintentionally put less effort into the initial, unprocessed reads.

“In this observer performance study of the visualization software, the readers demonstrated a statistically significant increase of 0.098 in the area under the LROC curve and a 16.8 percentage point increase in cancer detection sensitivity. This was accompanied by a 4.3 percentage point decrease in specificity. These findings indicate that the addition of the software-enhanced image should be of important clinical utility,” Freedman and co-authors concluded.