Thorax imaging fulfills promise with computer-aided detection

No technological advance for imaging the thorax has had so profound an impact on the visualization and management of data as multislice CT. Its fast acquisition times and higher resolution have forever changed imaging of the lung, an organ whose dynamic nature demands rapid acquisition and whose complex anatomy demands isotropic resolution for fine anatomic detail and true 3D analysis. MSCT allows shorter breath-holds, fewer breathing and motion artifacts, and reduced contrast delivery for CT angiography studies. The improved anatomic detail clarifies pulmonary anatomy and detects change as an indicator of disease.

The rapid clinical acceptance of MSCT scanners is an indication of their significant technological advances. The uptake of MSCT scanners in the U.S. between 1998 and 2001 had a compound annual growth rate of 160% in an otherwise dormant CT market. This dramatic increase reflects the increased clinical value of the faster, higher resolution scanners. The number and complexity of procedures now possible with MSCT will continue to increase as new protocols emerge that take advantage of its unique properties.

The finer resolution of data produced by MSCT, however, has the consequence of a massive explosion of data. The enormity of the acquired image volume has resulted in substantial workflow challenges to radiologists accustomed to diagnosing from axial slices, and it comes at a time when radiology departments are increasingly resource-constrained. Availability of radiologists increased just 3% in 2000, while CT procedures increased 19.4% in the same year, second only to PET, at 48%, in procedural increase.¹ The number of radiology-related job seekers has declined since the mid '90s, while the number of available jobs has steadily increased.

Because of expanded case demand and decreased resources, radiologists are flooded with information and desperately require better interpretive tools to analyze it quickly and accurately. Soft-copy review, more prevalent in CT than in any other modality, has already been shown to increase radiologists' productivity compared with film reading.² Soft-copy review is a minimal requirement with MSCT. The number of axial slices now generated with MSCT precludes their efficient review using 2D viewing techniques alone, and because of sheer numbers of images, they are often read at a lower resolution.

Volumetric postprocessing tools, available for over a decade,^3-8 are readily attainable in a soft-copy environment but are rarely used for primary diagnosis.⁹ This is in part because current volumetric analysis tools are not as efficient and accurate as axial review.¹⁰ Although acquired volumetric image resolution is accessible at the expense of many transverse images, true volumetric reconstruction with better postprocessing tools is needed to help the radiologist.

The long-awaited shift from 2D to 3D review¹¹ may be imminent, spearheaded by the rapid market acceptance of MSCT scanners and the subsequent demand for computer assistance to interpret these multidimensional exams.

CAD IN MEDICAL IMAGING

One measure of effective computer-assisted detection in radiological exams is its ability to focus valuable radiologists' time on interpretation of images. With the increasing volumes of single- and multimodality data, radiologists will have little time to spend on each examination and will have to focus their attention on interpretation. CAD has the potential to minimize their involvement in areas that do not require their expertise, such as image registration and automatic measurement. CAD can accomplish this by directing the radiologist's review and automatically generating quantitative information about the finding for faster interpretation.

Another measure of the effectiveness of CAD is its consistency and repeatability in identifying patterns that indicate abnormalities. Combined with the variability of human observation, this consistency can increase radiologists' accuracy. Readers of different experience levels will be affected differently.^12,13 Volumetric imaging is essential with MSCT acquisitions of the thorax, and CAD brings added clinical utility to volumetric imaging that warrants its more routine use in a diagnostic setting.

In their current form, CAD systems are used for automatic detection in medical images and as a second review for radiologists. In addition to improved accuracy, they provide other analysis tools that aid clinical interpretation. Given a consistent set of patterns, CAD systems can identify potential abnormalities, but the diagnosis of the finding remains the responsibility of the radiologist.

A radiologist reads the case and queries the CAD system, which highlights areas that may require a second review. The radiologist can react to the CAD mark if it was overlooked in the original interpretation or dismiss it altogether. As a second review, CAD can complement the inherent variability of human observation by finding oversights in the original review. The CAD system does not tire, take phone calls, change moods, or need to rush home for dinner. It may not need the sensitivity of the radiologist to be valuable,¹⁴ but the consistency of detection can help the radiologist.¹⁵

- CAD for MSCT of thorax. Potential CAD solutions to the data management problem are quickly moving from research to marketplace. Automatic detection may provide a key component to the workflow difficulties facing current MSCT users.

CAD systems are being prototypically evaluated for the automatic detection of lung nodules using MSCT,¹⁶ but unlike mammography and projection chest exams, no CAD systems for this application are currently available. CAD can target patterns indicative of lung nodules after an initial clinical evaluation, automatically measure lesion size and density, and make comparisons over time. This combination of tools provides a more efficient review for the radiologist.

Most CAD systems under evaluation concentrate on lung nodule detection,^12,17-21 and more recent work has focused on CAD for pulmonary embolism (PE) detection.^13,18,19,22 Both disease states lend themselves well to the help of CAD because they have proven difficult to detect and involve anatomic complexities and image volumes that could potentially increase reading time with MSCT using transverse sections.

- CAD for MSCT lung nodule detection. Traditional CT and single-detector spiral CT have been shown to be significantly better at detecting and diagnosing focal lung disease than chest radiographs.^23,24 But even with the increased anatomic clarity of the lung using CT, cancers are still missed.^25-28 Earlier methods developed to automatically detect lung nodules using single-detector-row CT were composed largely of 2D analyses with some compensation for the three dimensionality of CT data sets.^29-33 With the transition to isotropic resolution possible with MSCT of the chest, these techniques may not scale to true 3D analysis.

The use of CAD for MSCT analysis of lung nodules and other disease states is less mature than its use in mammography because of the limited time that both MSCT and CAD systems for MSCT have been available. At the base of the CAD system are the automatic detection algorithms that identify regions consistent with suspicion of disease. Most reported techniques have demonstrated high sensitivity for lung nodule detection.^{13,17-22,32,34,35} Preliminary results indicate that CAD, with high sensitivity and low false-positive rates, could perform acceptably in routine clinical practice.

Stand-alone CAD for lung nodule detection, however, varies from its performance in mammography. A comparable standard of biopsy proof is not only difficult but may not apply to the detection of lung nodules. First, the biopsy of nodules as small as 3 or 4 mm in diameter is uncommon, and therefore a database of small cancers is difficult to acquire, but it is likely that smaller abnormalities pose the larger detection problem. Second, detection algorithms designed to identify lung nodules as an abnormality are not designed to detect malignancies, as not all lung nodules are malignant. They may, however, be important clinical findings. Biopsy proof, while objective, may not be a scalable reference criterion for lung nodule detection.

Generally, gold standard algorithms for the detection of lung nodules are defined by consensus panels^12,17,20,21 but by their very nature are a more subjective standard than biopsy. Obtaining a repeatable reference data set with consensus panels relies on a very tight definition of the reference measurement. But radiologists in test environments, removed from their normal clinical environment, may overcall lesions because of the lack of clinical consequence or patient impact of their decision making. All factors contribute to the difficulty in defining the reference data set.

One method of producing a repeatable reference data set is through the insertion of artificial nodules with properties similar to true nodules into the original imaging database.^{5,13,18,19,27} This tighter standard may prove valuable in the evaluation of stand-alone performance in CAD systems, but it has yet to be tested on a wider scale.

INTEGRATION INTO WORKFLOW

The integration of CAD into the current radiologic environment of MSCT is challenging, partly because MSCT has so notably affected the clinical workflow that it requires fundamental change.^36,37 CAD can facilitate that change: Common components of current systems display the detected nodule, its contours, and 2D and volumetric nodule measurements. These measurements can be saved and tracked over time to automatically evaluate nodule progression or regression. Once the nodule is automatically detected, these workflow enhancements, which are otherwise done manually in most cases, are available to the user. The automatically generated measurements could save time and produce more repeatable results than human interaction. CAD therefore has immediate potential to focus the radiologist's energy on interpretation of images and to increase efficiency in a MSCT environment.^{12,13,18,19,21,33,38}

The ability to retrospectively reconstruct axial slices at varying slice thicknesses is another workflow enhancement driven by MSCT and CAD.²⁰ CAD will always read the highest resolution image available, but because it remains routine to read MSCT data using transverse reconstructions, the user may choose to read fewer slices at a lower resolution. Ideally, CAD could influence the radiologist's interpretation if lung nodules are overlooked in a normal clinical setting. Wood et al¹⁷ reported a 7.3% increase in the detection of actionable lung nodules after double reading of MSCT data sets, indicating the potential for a higher rate of detection from CAD after just a single read.

The value of CAD for increased accuracy of detection is found in the clinically important findings not originally detected with the double read. Das et al¹² reported the effectiveness of CAD for a reader with six years' experience reading chest CT images and another reader with one year of experience, before and after the use of the CAD system. With CAD, the performance of the inexperienced reader approached that of the experienced reader. CAD therefore has the potential to level the playing field among users with various levels of experience, producing a more consistent performance.

The observer performance of CAD for detection of lung nodules is not sufficiently developed to determine its effect on clinical outcome.

PULMONARY EMBOLISM

MSCT has made the greatest impact in vascular applications. MSCT for PE diagnosis can potentially reestablish the gold standard from more expensive and invasive methods such as traditional angiography. CAD is ideally suited for the detection of PE because of the tortuous nature of the pulmonary arterial tree.

Tracking its branching structure to identify emboli is time-consuming and tedious, especially with the more common mode of reviewing CTA examinations with transverse slices. Once an abnormality is found, however, its diagnosis is reasonably straightforward. Identifying PE in CTA examinations is a problem of detection, making it an ideal use of CAD, which can potentially reduce tracking time, increase accuracy of detection, and provide a valuable safety net for physicians making an emergent and critical diagnosis. CAD with MSCT makes this application clinically feasible in a resource-constrained environment.

CAD has demonstrated a high level of performance in initial studies that approach clinical acceptability. Combined with CTA studies using MSCT, CAD could provide a more accurate and less invasive clinical evaluation for PE than traditional angiography. Algorithms have been developed that combine PE detection programs with lung nodule detection systems.

A lung nodule found in a PE study, for example, would demarcate the regions suspicious for a lung nodule even though those nodules may not be visible in the PE review window. The automatic detection of both PE and lung nodules is available in the same application, which further optimizes the clinical workflow in an MSCT environment.

OTHER APPLICATIONS

CAD applications focus on the detection of abnormalities of the thorax rather than the diagnosis of these abnormalities, but as these systems become more advanced, their output will likely be of higher diagnostic quality. The transition of these devices from detection to diagnosis will probably be led by mammography³⁹ because of the longevity of mammographic systems in clinical practice and their familiarity to regulatory agencies. A computer-aided diagnosis product would probably be commercially acceptable after its use as a detection product has been proven.

Similarly, CAD has the potential to be a first reader in some clinical applications, and the sensitivity of computerized detection may surpass human observation for certain disease states, such as calcifications.

Computerized detection is currently being used as a first reader for Pap smears and may in time become the first reader in screening environments where the radiologist reads predominately normal cases. These first reader applications would again be preceded by clinical acceptance as a second reader.

Because the workflow in the current MSCT environment is changing, CAD could have value as a concurrent reader, with the CAD marks available to the radiologist as he or she first reviews the case.

CAD's acceptance as a second review, however, would have to be well established before it approaches the standard of a concurrent reader. Automatic detection of abnormalities of the thorax could in the future be combined with surgical navigation, as both technologies are independently well established. Similarly, the integration of CAD information with data from anatomic (MSCT) and functional (PET) imaging will play a key role in supporting the unification and workflow of these two technologies.

Present-day CAD systems complement human observation by consistency and repeatability without fatigue and therefore augment the natural variability of human observation. In their present form, however, they lack the judgment of the human observer; therefore, it remains the radiologist's responsibility to determine the patient's clinical course.

Given the dynamic nature and anatomic complexity of the organs in the thorax, faster and higher resolution acquisitions substantially affect the quality of the examination. MSCT has the potential to fundamentally change the way CT of the thorax is performed if the workflow problems associated with larger data sets can be resolved in the future.

With the known miss rates of human observations added to the complexity of MSCT image overload, observational oversights could become more prevalent with higher resolution data. Workflow issues associated with MSCT may thus require CAD in a clinical setting to increase accuracy of detection, assuredness and consistency of review, and productivity. These improvements may enable volumetric imaging to reach its true clinical potential.

References

1. Margulis A. International Society of Strategic Studies in Radiology (ISSSR), San Francisco, CA, 2001.

2. Reiner BI, Siegel EL, Hooper FJ, et al. Radiologists' productivity in the interpretation of CT scans: a comparison of PACS with conventional film. AJR 2001;176:861-864.

3. Napel S, Marks MP, Rubin GD, et al. CT angiography with spiral CT and maximum intensity projection. Radiology 1992;185:607-610.

4. Napel S, Rubin GD, Jeffrey RB Jr. STS-MIP: a new reconstruction technique for CT of the chest. JCAT 1993;17:832-838.

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

6. Johnson PT, Fishman EK, Duckwall JR, et al. Interactive three-dimensional volume rendering of spiral CT data: current applications in the thorax. Radiographics 1998;18:165-187.

7. Fan L, Qian J, Wei G-Q, Naidich DP. Pulmonary nodule detection using cartwheel projection analysis. Presented at CARS. Berlin, Heidelberg, New York: Springer; 2002.

8. Gruden JF, Ouanounou S, Tigges S, et al. Incremental benefit of maximum-intensity-projection images on observer detection of small pulmonary nodules revealed by multidetector CT. AJR 2002;179:149-157.

9. Fishman EK. International Society for Strategic Studies in Radiology (ISSSR), San Francisco, CA, 2001.

10. Rubin GD. 3-D imaging with MDCT. Eur J Radiol 2003;45(suppl 1):S37-S41.

11. Rubin GD, Napel S, Leung AN. Volumetric analysis of volumetric data: achieving a paradigm shift. Radiology 1996;200:312-317.

12. Das M, Jacobson F, Schneider A, et al. CAD of lung nodules on multidetector-row CT: impact on readers with variable levels of experience. Radiology 2002;225:476.

13. Das M, Schneider A, Jacobson F, et al. CAD of lung nodules on multidetector-row CT: impact on readers with variable levels of experience. Eur Radiol 2003;13:315.

14. Giger M, MacMahon H. Image processing and computer aided diagnosis. Radiol Clin N Am 1996;34:565-596.

15. Warren Burhenne LJ, Wood SA, D'Orsi CJ, et al. Potential contribution of computer-aided detection to the sensitivity of screening mammography. Radiology 2000;215:554-562.

16. Lawler L, Wood S, Fishman E. Computer-assisted detection of pulmonary nodules: preliminary observations using a prototype system with multidetector row CT datasets. J Digit Imaging 2003 Sept;16(3):251-261.

17. Wood S, Stapleton S, Schneider A, et al. CAD of actionable lung nodules using MSCT scans of the lung: sensitivity and false marker rates. Radiology 2003;225:477.

18. Das M, Schneider A, Anderson M, et al. Computer-aided detection (CAD) of lung nodules in high resolution multidetector-row CT (MDCT) chest scans. Eur Radiol 2003;13:169.

19. Das M, Schneider A, Bonneau V, et al. CAD of segmental and subsegmental pulmonary emboli on 1-mm multidetector-row CT studies. Eur Radiol 2003;13:138.

20. Wood S, Schneider A, Gamboa-Aldeco A, Castellino R. The ImageChecker-CT: computer-aided detection (CAD) for chest MSCT. Eur Radiol 2003;13:574.

21. Wood S, Stapleton S, Schneider A, et al. CAD of actionable lung nodules using MSCT scans of the lung: sensitivity and false marker rates. J Eur Radiol 2003;13:169.

22. Schoepf U, Das M, Schneider A, et al. CAD of segmental and subsegmental pulmonary emboli on 1-mm multidetector-row CT studies. Radiology 2003;225:384.

23. Muhm JR, Brown LR, Crowe JK. Use of computed tomography in the detection of pulmonary nodules. Mayo Clin Proc 1977;52:345-348.

24. Muhm JR, Miller WE, Fontana RS, et al. Lung cancer detected during a screening program using four-month chest radiographs. Radiology 1983;148:609-615.

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

26. Hazelrigg SR, Boley TM, Weber D, et al. Incidence of lung nodules found in patients undergoing lung volume reduction. Ann Thorac Surg 1997;64:303-306.

27. Rusinek H, Naidich DP, McGuinness G, et al. Pulmonary nodule detection: low-dose versus conventional CT. Radiology 1997;209:243-249.

28. Kakinuma R, Ohmatsu H, Kaneko M, et al. Detection failures in spiral CT screening for lung cancer: analysis of CT findings. Radiology 1999;212:61-66.

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

30. 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.

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

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

33. 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.

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

35. Brown MS, Goldin JG, Suh RD, et al. Lung micronodules: automated method for detection at thin-section CT-initial experience. Radiology 2003;226:256-262.

36. Rubin GD. 3-D imaging with MDCT. Eur J Radiol 2003;45(suppl 1):S37-S41.

37. Rubin GD. MDCT imaging of the aorta and peripheral vessels. Eur J Radiol 2003;45(suppl 1):S42-S49.

38. Fan L, Qian J, Wei G-Q, Naidich DP. Pulmonary nodule detection using cartwheel projection analysis. Presented at CARS, Berlin, Heidelberg, New York: Springer; 2002.

39. Giger M, Huo Z, Vyborney C, et al. Results of an observer study with an intelligent mammographic workstation for CAD. In: Peitgen H-O, ed. Digital mammography. Presented at International Workshop on Digital Mammography. Berlin, Heidelberg, New York: Springer; 2002:297-303.

Dr. Schoepf is an associate professor of radiology and director of CT research and development at the Medical University of South Carolina. He has received grants/research support from Bracco, GE Healthcare, Medrad, and Siemens and is a consultant to Bracco, GE Healthcare, PAION, Siemens, and Voxar.