Computer-Aided Screening Detects Missed Lung Cancers

March 1, 2002
Oncology NEWS International, Oncology NEWS International Vol 11 No 3, Volume 11, Issue 3

CHICAGO-Screening for lung cancer with low-dose helical CT scans is becoming increasingly popular. Computer programs to assist in the detection of lung cancers appear to increase the accuracy of CT screening, said Samuel G. Armato III, PhD, assistant professor of radiology, University of Chicago.

CHICAGO—Screening for lung cancer with low-dose helical CT scans is becoming increasingly popular. Computer programs to assist in the detection of lung cancers appear to increase the accuracy of CT screening, said Samuel G. Armato III, PhD, assistant professor of radiology, University of Chicago.

In a study he reported at the 87th Scientific Assembly and Annual Meeting of the Radiological Society of North America (RSNA abstract 380), radiologists missed 38 biopsy-confirmed lung cancers out of a total of 50 lung nodules during patients’ initial clinical workup. When a computerized detection program was applied, 32 (84%) of the previously missed cancers were found.

Dr. Armato explained that research involving the computerized lung cancer detection system, which was developed at the University of Chicago, is still at the experimental stage. "Our research at this point was not intended to show how radiologists would use the technology clinically but to show how well the computer itself performed in detecting missed lung cancer," he said.

Especially important, he added, was that the study targeted a set of problematic clinical images. "This was a difficult database, by definition, because these are all cancers that were missed by radiologists. So a radiologist, in the category of detection errors, was batting zero percent and our computer program was batting 78%. That’s a pretty good sensitivity for this kind of situation," he said.

The computerized lung nodule detection system segments the lung images to create a segmented lung volume, then thresholds the lung volume and identifies three-dimensional contiguous structures in each thresholded lung volume data set.

The system selects lesions as possible nodules on the basis of size criteria. A rule-based approach is used to decrease the number of false-positive nodules. An automated program distinguishes between true lesions and normal anatomy in the remaining nodule candidates.

In the study, Dr. Armato categorized the 38 missed cancers as either radiologist detection errors (23 cancers) or interpretation mistakes (15 cancers). The computer program found 18 (78%) of the 23 detection errors and 14 (93%) of the 15 misinterpreted lesions. In both categories, the computerized program had a 1.6 false-positive rate per CT section image.

In addition to focusing on reducing the number of false positives, Dr. Armato and his associates will be testing the system as part of standard radiology practice. "The next major step in this research overall will be to show that these methods, when used in conjunction with the radiological decision-making process, actually improve radiologists’ performance," he said. 

Related Content:

Lung Cancer