Heidi C. Roberts, MD, FRCP©

CAD no longer stands merely for computer-aided detection. Today's robust software programs make it possible to help characterize, or diagnose, nodules, particularly in the lungs. Current computer-aided technology favors detection, but the ability to diagnose lung nodules based on certain characteristics continues to develop rapidly. It's only a matter of time before computer-aided detection and computer-aided diagnosis become mainstay tools for pulmonary evaluation.

The development of CAD tools for lung nodule evaluation seems to parallel the application of lung cancer screening using low-dose regimens. Newer scanners with multidetector-row capabilities use screening protocols to scan a whole lung in one breath-hold with thin (1 to 1.25-mm) slices.

 Suddenly, radiologists are faced with an overwhelming number of images that need to be searched for the presence of pulmonary nodules.

Even though almost all parenchymal nodules found in a screening population are benign,

 it is important to find all of them, since any one could be a cancer. Several studies have emphasized that lung cancers are frequently missed on routine or screening CT and can be seen retrospectively.

 Subsequent studies demonstrated that CAD found a large proportion of these missed cancers.

One major expected and documented benefit of CAD is improved diagnostic accuracy. Commonly, size thresholds for lung nodule detection are incorporated into the software algorithm, which can be fixed or more flexibly defined by the user. While early screening protocols targeted nodules regardless of their size,

 nodules smaller than 5 mm are unlikely to be of relevance.

 While this size recommendation has become widely accepted, its exact implementation differs in screening programs that use a slightly different cutoff for actionable nodules. Generally, this lies between 4 and 6 mm.

 Applying a size filter to discard small nodules makes CAD more appealing, since it reduces the number of false-positive results considerably.

In general, CAD software uses round or triangular markers, presented on the screen and overlying the original CT image, to surround a possible nodule for further review (Figure 1). Assuming that CAD is used to catch missed lung nodules and, thus, potential cancers, all nodules above a given size threshold must be labeled by the software.

A true-positive CAD mark is any parenchymal nodule, whether benign or malignant (Figure 2). The performance marker describing how well CAD finds nodules is sensitivity (the number of true-positive results divided by the sum of true-positive and false-negative, or missed, nodules).

 Sensitivity does not necessarily need to be high, as long as CAD finds nodules missed by the radiologist. Indeed, studies have shown that CAD does find different nodules from a radiology reader.

 Unfortunately, direct comparisons of CAD for lung nodule detection are rather limited, as systems are evaluated in different patient populations, on CT scans with different data acquisition parameters, and with software of different manufacturers. Studies report sensitivities of stand-alone CAD between 70% and 97%

 and unequivocally report improved sensitivities when CAD is added to the radiologist's reading.

Whether CAD is able to decrease the time for reading a chest CT depends largely on its integration into the workflow. CAD needs to be integrated into an existing RIS/PACS, rather than being available on a separate dedicated workstation. R2 Technology (Sunnyvale, CA), for example, has taken a first step. Its CAD results can be accessed across a PACS network via DICOM secondary capture or gray-scale presentation state (GSPS) from the R2 server.

The timing of CAD usage is also important. CAD can be used as a second or concurrent reader. Most systems are promoted as a second reader to be used after the radiologist has already scrutinized the case. This increases the reading time and targets diagnostic accuracy more than improvement of workflow. A concurrent read model, used simultaneously with initial reading of a CT study, does not impair workflow in the same manner. This approach, as implemented in the LungCAR from Medicsight (London) however, demands a very high sensitivity. If radiologists instantly have CAD marks available, they invariably decrease their attention and come to rely on the software to find nodules.

Another important CAD performance parameter is the number of false positives. False-positive marks are defined as any CAD mark in an anatomic or artifactual structure that is not a nodule, such as artifacts from respiratory or cardiac motion, vessel bifurcations, hilar vessels, parenchymal scars, or normal mediastinal organs, osteophytes, etc. (Figure 3).

 Even though most false positives are easily dismissible, simply by scrolling up and down a few slices,

 the rejection of a high number of individual marks does take time.

Good CAD performance is generally defined as high sensitivity combined with a low number of false positives. The trade-off between time and diagnostic accuracy is tolerated differently by individuals, however. The acceptable sensitivity/false-positive ratio is thus not a fixed parameter but one that dynamically changes with the reader and the database under evaluation. A junior or inexperienced reader, for example, might require higher CAD sensitivity, since he or she is much more likely to miss nodules, and might also be more tolerant of the additional reading time required to reject false positives. The incremental benefit of CAD has been shown to decrease with the experience of a radiologist.

Similarly, sensitivity needs to be higher if CAD is applied to an oncology population, where all nodules, even small ones, are likely significant and the size threshold cannot be used as a filter.

 Both CAD and radiologist readers have lower detection rates with decreasing nodule size, but an overall incremental improvement in diagnostic accuracy occurs even when CAD is used to detect small nodules.

CAD software is hardly ever used as an adjudicator for questionable nodules and thus diagnostic purposes.

 Tools are available within most software packages, however, that help further characterize a lung nodule once found. This additional information is gained from density and-in the case of follow-up studies-3D volume measurements and computation of growth rate and doubling time.

The presence and percentage of calcification is displayed on the screen. This may be helpful in completely or almost completely calcified nodules, but the morphologic pattern of calcification is more important than the mere presence of calcification itself.

 Reader interpretation required to assess such patterns cannot yet be replaced.

More informative are CAD comparisons when CT examinations at different time points are available. The 3D display of lung nodules with volume measurements is a common CAD feature, with automated computation of interval change/growth and doubling time (Figure 4). Doubling times allow inference of the nature of a nodule: Benign nodules are generally thought to have a volume doubling time of more than 400 days, which is equivalent to the absence of growth in a lesion over a two-year period.

 This dogmatic postulation, which stems from chest radiography studies of the 1970s, has recently been weakened and more realistically has a predictive value of only 67%.

 Similarly, the typical doubling time of lung cancer is currently being redefined. Initially reported as averaging 163.7 days,

 slower growing adenocarcinomas in particular found in the screening population have a longer volume doubling time, with a mean of 452 plus/minus 381 days (ranging from 52 to 1733 days).

Theoretically, more subtle changes in nodule size can be detected when volume, rather than diameter, is the target measurement. Actual doubling of a sphere volume implies a diameter change of only 26%, which can easily be overlooked, particularly in smaller nodules.

Under satisfactory technical requirements, computer-assisted volumetry for solid nodules is accurate, robust, repeatable, and consistent.

 Unfortunately, in real life, technical requirements are far from perfect, and nodules are not always completely solid and well defined, resulting in limited reproducibility of nodule volumes.

 Validation of volume comparison in different scenarios is under way, and it may define a threshold of apparent growth that can be reliably used to identify a malignant nodule.

Volume measurements may vary through differences in data acquisition (mAs, collimation, field-of-view), patient-related factors (different breath-hold levels, motion), software (inclusion/exclusion of adjacent structures such as vessels or pleura), and the nodule itself (density, inclusion or exclusion of surrounding ground glass or spiculations).

CAD for lung nodule detection is extensively developed and can be recommended for use as a concurrent or second reader to find nodules in thin-slice CT images. This recommendation more or less limits the CAD indication to lung cancer screening protocols. Routine examinations, such as oncology baseline or follow-up CT, are commonly obtained with larger slice thicknesses, in the magnitude of 5 mm. The decrease in performance of CAD with increasing slice thickness has been documented for 5 to 10-mm increments,

 and it is expected also in 1 to 5-mm increments but has not yet been documented in a large study. Some manufacturers restrict their algorithm to process only thin-slice CT images 2 mm or smaller. The benefit of CAD for 5-mm slices or larger is not yet known.

Growth assessments can reliably be used for nodule characterization, if the CT scans have been performed with identical FOV, collimation, dose, and-as far as assessable-breath-hold.

Remaining issues lie with the nodules themselves. Solid or even part-solid nodules are more easily detected by CAD, whereas the detection of pure ground-glass opacities (GGO) remains the radiologists' responsibility. Some manufacturers limit their detection to solid nodules only, while others are targeting GGOs as well. Studies are under way to evaluate the performance of CAD systems in the detection of GGOs.

From the first major representation of CAD for lung nodule detection at the RSNA meeting in 2001 to today, we have witnessed a rapid development that has not yet plateaued. Studies are published almost monthly, evaluating CAD performances and addressing its limitations.

1. International Early Lung Cancer Action Program. www.ielcap.org.

2. Bepler G, Goodridge Carney D, Djulbegovic B, et al. A systematic review and lessons learned from early lung cancer detection trials using low-dose computed tomography of the chest. Cancer Control 2003;10:306-314.

3. Gurney JW. Missed lung cancer at CT: imaging findings in nine patients. Radiology 1996;199:117-122.

4. Li F, Sone S, Abe H, et al. Lung cancers missed at low-dose helical CT Screening in a general population: comparison of clinical, histopathologic, and imaging findings. Radiology 2002;225:673-683.

5. Armato SG, Li F, Giger ML, MacMahon H, et al. Lung cancer: Performance of automated lung nodule detection applied to cancers missed in a CT screening program. Radiology 2002;225:685-692.

6. Li F, Arimura H, Suzuki K, et al. Computer-aided detection of peripheral lung cancers missed at CT: ROC analyses without and with localization. Radiology 2005;237:684-690.

7. Henschke CI, McCauley DI, Yankelevitz DF, et al. Early Lung Cancer Action Project: overall design and findings from baseline screening. Lancet 1999;354:99-105.

8. Swenson SJ, Jett JR, Hartman TE, et al: CT screening for lung cancer: five-year prospective experience. Radiology 2005;235:259-265.

9. Henschke CI, Yankelevitz DF, Naidich DP, et al. CT Screening for lung cancer: suspiciousness of nodules according to size on baseline scans. Radiology 2004;231:164-168.

10. National Lung Screening Trial Protocol. www.acrin-nlst.org.

11. Gohagan J, Marcus P, Fagerstrom R, Pinsky P et al. Baseline findings of a randomized feasibility trial of lung cancer screening with spiral CT scan vs chest radiography: The Lung Screening Study of the National Cancer Institute. Chest 2004;126:114-121.

12. Pastorino U, Bellomi M, Landoni C, et al. Early lung-cancer detection with spiral CT and positron emission tomography in heavy smokers: 2-year results. Lancet 2003;362:593-597.

13. Giger ML, Bae KT, MacMahon H. Computerized detection of pulmonary nodules in computed tomography images. Invest Radiol 1994;29:459-465.

14. Croisille P, Souto M, Cova M, et al. Pulmonary nodules: improved detection with vascular segmentation and extraction with spiral CT. Radiology 1995;197:397-401.

15. Kanazawa K, Kawata Y, Niki N, et al. Computer-aided diagnosis for pulmonary nodules based on helical CT images. Comput Med Imaging Graph 1998;22:157-167.

16. Armato SG, Giger ML, Moran CJ, et al. Computerized detection of pulmonary nodules on CT scans. Radiographics 1999;19:1303-1311.

17. Reeves AP, Kostis WJ. Computed-aided diagnosis of small pulmonary nodules. Semin Ultrasound CT MR 2000;21:116-128.

18. Bae KT, Kim JS, Na YH, et al. Pulmonary nodules: Automated detection on CT images with morphologic matching algorithm-preliminary results. Radiology 2005;236:286-294.

19. Lee JW, Goo JM, Lee HJ, et al. The potential contribution of a computer-aided detection system for lung nodule detection in multidetector row computed tomography. Invest Radiol 2004;39:649-655.

20. Roberts HC, Patsios D, Kucharczyk M, et al. The utility of computer-aided detection (CAD) for lung cancer screening using low-dose CT. International Congress Series 2005;1281:1137-1142.

21. Naidich D, White C, Meziane M, et al. A Multicenter, Multi-reader retrospective clinical study to assess the incremental value of LUNG CAD in the identification of pulmonary nodules on thoracic CT examinations. 91st Annual Meeting RSNA, Chicago, 2005.

22. Brown MS, Goldin JG, Rogers S, et al. Computer-aided Lung Nodule Detection in CT: Results of Large-Scale Observer Test. Acad Radiol 2005;12:681-686.

23. Awai K, Murao K, Ozawa A, et al. Pulmonary nodules at chest CT: Effect of computer-aided diagnosis on radiologists' detection performance. Radiology 2004;230:347-352.

24. Peldschus K, Herzog P, Wood SA, et al. Computer-aided diagnosis as a second reader: spectrum of findings in CT studies of the chest interpreted as normal. Chest 2005;128:1517-1523.

25. Beyer F, Fallenberg E, Juergens K. Computer-aided detection (CAD) of pulmonary nodules at MDCT-comparison of sensitivity and reading time for CAD as concurrent and second reader. 91st Annual Meeting RSNA, Chicago, 2005.

Heidi C. Roberts, MD, FRCP©

Articles

CAD for pulmonary evaluation approaches mainstay status