Chest Gains

In reviewing lung image data sets from multidetector row CT (MDCT), chest radiologists have employed two versions of automated analysis: computer-aided detection and computer-aided diagnosis. They possess the same acronym-CAD-but very different algorithms. Of these two forms of analysis, the most advanced indication for analyzing lung image data sets is detecting pulmonary nodules. This capability preceded the diagnostic differentiation of lung nodules as benign or malignant.

In recent years, researchers and manufacturers have developed CAD algorithms to detect and quantify diffuse lung diseases-both those associated with increased density, such as interstitial lung disease, and those linked with decreased density, such as emphysema.

CAD for nodules

Lung nodules represent CAD's first and most advanced chest application for detection and diagnosis using volume measurements, temporal comparison and morphological analysis. Several manufacturers offer this application; each system displays CT images either with or without a marker/prompt that highlights a potential nodule, either side by side or with an on/off option during scrolling. The radiologist can accept or reject those marks as nodules and add nodules that the system didn't find to produce a composite report.

Lung nodule detection has its market in the context of lung cancer screening and oncology cohorts--two high-volume areas that require high sensitivity while maintaining a large throughput.

Nodule detection

An ideal CAD performance in this context would require detection of all nodules--i.e., 100 percent sensitivity-without false positives. Radiologists can use one of three strategies for lung nodule detection via CAD.¹

In the first strategy, CAD serves as a second reader--i.e., a "spell-checker"-that increases the sensitivity of the first reader (radiologist) but also increases reading time, thereby reducing workflow. Most CAD systems are used in this manner.²

In the second strategy, CAD is a concurrent reader. This strategy requires a high sensitivity, as the radiologist invariably wouldn't examine unmarked areas closely-yet this practice would increase reading time. Few CAD systems are developed as concurrent readers.³

In the third strategy, CAD is the first reader. This scenario arises predominantly from lung cancer screening studies, which scan large numbers of images for tiny nodules. This application isn't yet possible, however, as prereading would put the radiologist in the "spell-checker" position and require CAD to approach 100 percent sensitivity. To date, no CAD system has this sensitivity. Table 1 and Figure 1, which plot the nodule detection rate against the publication year, show a steady-but-slow increase in sensitivity over time.

CAD has maximum impact when it's used as a second reader, since CAD and radiologists tend to see different nodules; their combined detection rates achieve an overall sensitivity exceeding 90 percent.^4-7 Sensitivity depends on slice thickness (74 percent in thick slices vs. 80 percent in thin slices); nodule dimension (sensitivity increases with nodule size); nodule consistency (solid nodule sensitivity is roughly 80 percent vs. 73 percent in other nodules) and nodule positioning (CAD is better in central nodules; a radiologist performs better with subpleural and peripheral nodules).^4,23

Technical requirements for CT images in the successful application of CAD include a high in-plane resolution, narrow collimation and reconstruction intervals, and high-dose rather than low-dose scans. This approach best discriminates branching vessels and nodules.

In addition to sensitivity, CAD's other important performance measure for nodule detection is its number of false positives. While radiologists often dismiss these vessels, fissures, areas of atelectases or lung collapse, and artifacts from respiratory or cardiac motion  after scrolling a few images, they dramatically impact reading time and workflow, thus limiting general acceptance of CAD.

Fortunately, better software algorithms have reduced false positives significantly. In Table 1, the average number of false positives per patient was 30.5 in publications from 2001-04, and 3.7 per patient in publications from 2005-06.