Lung cancer has for some time ranked as the leading cause of cancer death in the U.S. It will account for 31% of cancer deaths among men and 25% of cancer deaths among women in this country in 2003, with an estimated total of 157,200 deaths.1 Although the scientific jury is still out regarding the impact of earlier lung cancer detection on patient mortality, improvements in diagnostic imaging technology continue to make earlier detection of lung cancer possible-and intuitively desirable.

A diagnostic task in radiology generally comprises two components: the detection task and the classification task. First, a radiologist must identify the location of an abnormality. The abnormality of interest here is the lung nodule. Second, once the presence of a nodule has been determined, the radiologist must assess its severity and recommend an appropriate course of action: Is the likelihood that the nodule represents a malignancy so high that a biopsy should be performed, or is a wait-and-watch approach more prudent? Research in computer-aided diagnosis has followed this diagnostic model of detection and classification.

CAD has been defined as a diagnosis made by a radiologist when the output of computerized image analysis methods has been incorporated into his or her medical decision-making process.2 Although some people distinguish between the concepts of computer-aided diagnosis and computer-aided detection, CAD may be interpreted broadly to incorporate both the detection and the classification tasks that radiologists face. The term "computer-aided diagnosis" refers more commonly to the actual computer methods that provide this diagnostic assistance, in the context of "computers as a diagnostic aid." Simply stated, CAD methods apply powerful techniques of physics, mathematics, statistics, and computer science to the anatomical and physiological information contained within medical images. A variety of CAD methods have been developed for lung nodules in both projection radiographs of the chest and thoracic CT scans.

The chest radiograph, long the mainstay of radiology, often provides the first opportunity for a radiologist to diagnose a patient with lung cancer. A complicated anatomy combined with perceptual problems that accompany the projection of a three-dimensional object (the patient) into two dimensions (the image plane), however, makes identification of lung nodules a burdensome task for radiologists. Even more burdensome is the classification of nodules as benign or malignant based on the radiographic image. CAD techniques are being developed to address these challenges.

Much progress has been made since one of the first published reports of automated lung nodule detection in chest radiographs appeared in the literature in 1973.3 The FDA (www.fda.gov) approved the first CAD device for chest radiography in July 2001, granting premarket approval to Deus Technologies for its RapidScreen RS-2000, a CAD device "intended to identify and mark regions of interest . . . on digitized frontal chest radiographs." Consistent with the concept of computer-aided diagnosis, however, this device was approved for use only after an initial interpretation of the radiograph by a radiologist. The evolution of the concept of "computer-aided diagnosis" to one of "computerized diagnosis" is still many years in the future.

CT DATA OVERLOAD

Since the introduction of the first clinical CT scanner in 1972, CT has played a continued and expanding role in radiology; CT of the thorax has been no exception. The development of helical CT scanning in the late 1980s boosted the clinical role of CT even further. Helical CT, which allows for continuous translation of the patient table during the acquisition of image data, was an alternative to the then-conventional "step-and-shoot" approach. Today, helical CT, with its many clinical advantages, is the conventional CT approach. Its superiority over chest radiography for detection of lung nodules has been demonstrated, and, in fact, helical CT is recognized as the most sensitive imaging modality for lung nodule detection.4 So great have been the successes of helical CT that radiologists are facing a new kind of challenge: information overload.

Large amounts of CT image data are being generated for radiologist interpretation, due to both the expanding diagnostic and screening roles of CT and the acquisition of a greater number of individual images per scan. Diagnostic studies that were once performed with other imaging modalities are now performed with CT. The ventilation/perfusion radionuclide scan for suspected pulmonary embolism, for example, is now largely a relic of the past; CT has become the preferred modality. Even diagnostic CT scans acquired for indications unrelated to cancer must be evaluated by radiologists for the possible presence of a lung nodule as an incidental finding.

Beyond its diagnostic role, helical CT (especially when combined with a low-dose imaging protocol) is gaining an international reputation as a viable tool for lung cancer screening. Several studies have found an increased number and a favorable stage shift of cancers detected at low-dose CT screening relative to screening with chest radiography and relative to the onset of clinical symptoms.5-7

Advances in CT imaging technology further expand radiologists' workload by generating a greater number of individual images per scan. A thoracic CT scan formerly produced approximately 30 sections with the 10-mm collimation that was standard for many years. The same type of scan, with the 1.25-mm collimation available on state-of-the-art multidetector scanners, now generates about 240 section images for radiologists to interpret. With an increase in the number of CT scans being performed for a wide variety of diagnostic and screening purposes compounded by an increasing number of images acquired during each scan, computerized techniques for the automated analysis of CT scans for disease (and especially for lung nodules that may represent lung cancer) are quickly becoming a necessity. These techniques are expected to provide valuable assistance to radiologists.

DETECTION METHODS

Researchers have developed a number of computerized lung nodule detection methods. One of the first entailed an automated nodule detection approach based on gray-level thresholding combined with analysis of the geometric characteristics of identified lung nodule candidates.8 Other researchers have followed with various gray-level thresholding-based methods,9-13 fuzzy clustering algorithms,14 spatial filtering,15 template-matching methods,16 object-based deformation procedures,17 morphological analysis,18 and model-based techniques.19,20

A CAD method developed at the University of Chicago9 considers the three-dimensionality of CT images and automatically analyzes structures within the scan to identify the structures that most likely represent lung nodules. This process begins with a section of a CT scan (Figure 1A) and automatically segments the lung regions (B) based on a gray-level threshold. The collection of lung segmentation regions for all sections in the scan is regarded as a segmented lung volume. Next, a series of gray-level thresholds is applied to the segmented lung volume to generate a set of 3D lung nodule candidates (C). Geometry- and brightness-based characteristics are used to reduce the number of false-positive structures within the set of nodule candidates. The nodule candidates that remain at this point (the detection output) may be presented as circles superimposed on the original CT images for radiologist review (D).

A recent study applied this method to a database of 38 low-dose CT scans (10-mm collimation, 10-mm reconstruction interval) from a lung cancer screening trial. An 84% nodule detection sensitivity with one false positive per section was reported (Figure 2).21 These nodules represented lung cancers "missed" by radiologists during the initial interpretation of the screening examinations. Another study that used this method reported 71% nodule detection sensitivity with 0.4 false positive per section using 38 diagnostic CT scans (7-mm collimation, 5-mm reconstruction interval).22

Automated nodule detection methods indicate the spatial location of suspected nodules but do not then attempt to estimate the likelihood that a detected nodule is malignant. Researchers have developed a number of computerized lung nodule classification methods, which typically begin with the manual or semiautomated extraction of a known lesion. These classification schemes are based on CT density matrices,23 model-based similarity measures,24 pattern classification using nodule density, size, shape, and texture features,25 co-occurrence matrices,26 and shape and surface curvature features.27,28

The automated detection and classification tasks have generally been regarded separately, an approach that represents one possible clinical application of CAD in which a radiologist, after recognizing a suspicious lesion in a CT scan, might select that lesion and have the computer generate an estimated likelihood of malignancy. Another clinical application of CAD will be a system that indicates both the location of (detects) and the likelihood of malignancy of (classifies) a lung nodule.29 Such a paradigm, in which the output of automated lung nodule detection is used as input to automated lung nodule classification, begins to approach the notion of fully automated analysis of lung nodules in CT scans.

The ability of a computer to detect or classify lung nodules accurately is a necessary but not sufficient component of CAD. Consistent with the term "computer-aided diagnosis," the objective of these methods is to assist radiologists in their decision-making process, so the presentation of CAD output to radiologists is an important consideration for the clinical acceptance of these technologies. The development of intelligent computer interfaces, such as those reported in mammography, remains an open area of exploration for lung cancer CAD.

RESEARCH EFFORTS

International research interest in CAD of the lungs, especially as it relates to lung cancer, has expanded dramatically in both academia and industry in recent years. As often occurs in CAD research, access to clinical images and reliable "truth" information becomes the rate-limiting step in the development of these techniques. Many researchers are not affiliated with medical centers, so access to clinical images is especially burdensome. Even for those who are associated with medical centers, the seemingly ready supply of clinical images that might be gathered for research is not without its barriers: The task of identifying and collecting appropriate images can be tedious, and necessary federal regulations that protect patient privacy rights place additional demands on the process. Recognizing these difficulties, while seeking to facilitate further advances in CAD for lung cancer, the National Cancer Institute initiated the Lung Image Database Consortium in 2001. Consisting of Cornell University, the University of California, Los Angeles, the University of Chicago, the University of Iowa, and the University of Michigan, the LIDC is making progress toward creation of a publicly available database resource for lung CAD researchers.

The current climate of technological advances and a greater demand for imaging studies of the chest pose two challenges for today's radiologist: a greater time commitment for the review of imaging studies to identify all abnormal lesions, and the accompanying responsibility to distinguish nodule features that likely differentiate malignant nodules from benign lesions. Continued research in CAD of the lung is expected to address both these radiologic challenges. As research evolves into development in this exciting field, radiologists and patients alike will benefit from the work of many dedicated scientists and engineers around the world.

References

1. Jemal A, Murray T, Samuels A, et al. Cancer statistics, 2003. CA Cancer J Clin 2003;53:5-26.

2. Giger ML. Current issues in CAD for mammography. In: Doi K, Giger ML, Nishikawa RM, Schmidt RA, eds. Digital mammography '96: Proceedings of the 3rd International Workshop on Digital Mammography. Amsterdam: Elsevier; 1996:53-59.

3. Ballard DH, Sklansky J. Tumor detection in radiographs. Comput Biomed Res 1973;6:299-321.

4. Remy-Jardin M, Remy J, Giraud F, Marquette C-H. Pulmonary nodules: detection with thick-section spiral CT versus conventional CT. Radiology 1993;187:513-520.

5. Sone S, Takashima S, Li F, et al. Mass screening for lung cancer with mobile spiral computed tomography scanner. Lancet 1998;351:1242-1245.

6. Kaneko M, Eguchi K, Ohmatsu H, et al. Peripheral lung cancer: screening and detection with low-dose spiral CT versus radiography. Radiology 1996;201:798-802.

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

9. Armato SG III, Giger ML, MacMahon H. Automated detection of lung nodules in CT scans: preliminary results. Med Phys 2001;28:1552-1561.

10. Ko JP, Betke M. Chest CT: automated nodule detection and assessment of change over time-preliminary experience. Radiology 2001;218:267-273.

11. Gurcan MN, Sahiner B, Petrick N, et al. Lung nodule detection on thoracic computed tomography images: preliminary evaluation of a computer-aided diagnosis system. Med Phys 2002;29:2552-2558.

12. Fan L, Novak CL, Qian J, et al. Automatic detection of lung nodules from multi-slice low-dose CT images. SPIE Proceedings 2001;4322:1828-1835.

13. Fiebich M, Wormanns D, Heindel W. Improvement of method for computer-assisted detection of pulmonary nodules in CT of the chest. SPIE Proceedings 2001;4322:702-709.

14. Kanazawa K, Kubo M, Niki N, et al. Computer assisted lung cancer diagnosis based on helical images. In: RT Chin, IP HHS, Naiman AC, Pong T-C, eds. Image analysis applications and computer graphics. Proceedings of the Third International Computer Science Conference. Berlin: Springer-Verlag, 1995:323-330.

15. Okumura T, Miwa T, Kako J, et al. Image processing for computer-aided diagnosis of lung cancer screening system by CT (LSCT). SPIE Proceedings 1998;3338:1314-1322.

16. Lee Y, Hara T, Fujita H, et al. Automated detection of pulmonary nodules in helical CT images based on an improved template-matching technique. IEEE Transactions on Medical Imaging 2001;20:595-604.

17. Lou S-L, Chang C-L, Lin K-P, Chen T-S. Object-based deformation technique for 3-D CT lung nodule detection. SPIE Proceedings 1999;3661:1544-1552.

18. Taguchi H, Kawata Y, Niki N, et al. Lung cancer detection based on helical CT images using curved surface morphology analysis. SPIE Proceedings 1999;3661:1307-1314.

19. Ryan WJ, Reed JE, Swensen SJ, Sheedy PF Jr. Automatic detection of pulmonary nodules in CT. In: Lemke HU, Vannier MW, Inamura K, Farman AG, eds. Computer Assisted Radiology (CARS '96). Amsterdam: Elsevier Science, 1996:385-389.

20. Brown MS, McNitt-Gray MF, Goldin JG, et al. Patient-specific models for lung nodule detection and surveillance in CT images. IEEE Trans Med Imaging 2001;20:1242-1250.

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

22. Armato SG III, Altman MB, LaRiviere PJ. Automated detection of lung nodules in CT scans: effect of image reconstruction algorithm. Med Phys 2003;30:461-472.

23. Cavouras D, Prassopoulos P, Pantelidis N. Image analysis methods for solitary pulmonary nodule characterization by computed tomography. Europ J Radiol 1992;14:169-172.

24. Henschke CI, Yankelevitz DF, Mateescu I, et al. Neural networks for the analysis of small pulmonary nodules. Clin Imaging 1997;21:390-399.

25. McNitt-Gray MF, Hart EM, Wyckoff N, et al. A pattern classification approach to characterizing solitary pulmonary nodules imaged on high resolution CT: preliminary results. Med Phys 1999;26:880-888.

26. McNitt-Gray MF, Wyckoff N, Sayre JW, et al. The effects of co-occurrence matrix based texture parameters on the classification of solitary pulmonary nodules imaged on computed tomography. Comput Med Imaging Graph 1999;23:339-348.

27. Kawata Y, Niki N, Ohmatsu H, et al. Quantitative analysis of internal texture for classification of pulmonary nodules in three-dimensional thoracic images. SPIE Proceedings 2000;3979:863-871.

28. Reeves AP, Kostis WJ, Yankelevitz DF, Henschke CI. Three-dimensional shape characterization of solitary pulmonary nodules from helical CT scans. In: Lemke HU, Vannier MW, Inamura K, Farman AG, eds. Computer Assisted Radiology (CARS '99). Amsterdam: Elsevier Science, 1999:83-87.

29. Armato SG III, Altman MB, Wilkie J, et al. Automated lung nodule classification following automated nodule detection on CT: a serial approach. Med Phys 2003;30:1188-1197.

30. Armato SG III. Update on the development of an automated lung nodule detection method for CT scans. In: Lemke HU, Vannier MW, Inamura K, et al, eds. Computer Assisted Radiology and Surgery (CARS 2002). Berlin: Springer; 2002:695-700.

DR. ARMATO is an assistant professor of radiology at the University of Chicago. He holds options to shares in R2 Technology.