Figure 1. A digital chest radiograph with CAD results indicating two suspicious findings: the arrow over the right upper lung correctly indicates a small cancer. The arrow on the left lower lung indicates a normal blood vessel, which is a typical false-positive result. Even this imperfect level of accuracy, CAD has been shown to improve radiologists’ diagnostic accuracy. Images courtesy of Heber MacMahon, MD, University of Chicago.

An assistant who is always attentive, whose keen eyes overlook nothing. No interobserver variation. Fewer biopsies for cancer screening that yield benign tissue.

Those were the goals when researchers began work on what was then called “computer-aided diagnosis.” The goals have seemed unrealistic, but software-based second readers are now proving their value in mammography and approaching general use in thoracic imaging. This article reviews the successes and potential of computer-aided detection (CAD).

The First Success: Mammography

The first CAD systems to win marketing approval from the US Food and Drug Administration (FDA) have been for diagnostic and screening mammography. Although CAD can be applied to digitized film-screen images, it is easier if one begins with digital images, and the benefits of CAD appear to be encouraging breast imaging centers to commit to digital mammography equipment.1

The reasons for using imaging to screen for breast-or any other-cancer are to locate all of the abnormalities (no false-negative studies) and to avoid identifying normal tissue as abnormal (no false-positive studies), finding all cancers while they are still curable. Several studies have confirmed the ability of CAD to improve the sensitivity of screening mammography (fewer false-negative studies) without an unacceptable increase in the recall rate. In view of the known ability of screening mammography to reduce deaths from breast cancer, it is not unreasonable to forecast that CAD will bring about a further decline in mortality from this disease.

Robyn L. Birdwell, MD, and associates at Stanford University provided an example of the improvement in sensitivity achievable with CAD.2 Those investigators reviewed 427 patients with biopsy-proved breast cancer for whom the most recent prior mammogram was available. Retrospective review revealed lesions on 286 of the earlier images. Further study of 110 of these patients showed 115 malignant lesions, which were characterized in detail. The CAD system identified 89 of the missed cancers, suggesting that CAD could have reduced the false-negative rate by 77%.

Among the prospective demonstrations of the value of CAD outside the university setting is the recent report by Timothy W. Freer, MD, and Michael J. Ulissey, MD, of the Women’s Diagnostic and Breast Health Center in Plano, Tex. In almost 13,000 screening mammograms over a period of 1 year, CAD produced a 19.5% increase in the number of cancers detected, with 78% of the cancers being discovered in stage 0 or stage I.3 Before the introduction of CAD, 73% of the breast cancers found on screening mammograms were of similarly low stage. Although more patients were recalled for further studies after CAD was introduced (7.7% vs 6.5%), this increase was considered acceptable. The positive predictive value for biopsy (the percentage of biopsies yielding cancer) was unchanged at 38%.

A hospital-based practice that has also demonstrated the value of CAD is the Joyce Martin Hampton Breast Center at Palmetto Richland Memorial Hospital in Columbia, SC, which performs 12,000 to 15,000 screening mammograms every year. A CAD system was installed there in January 1999, and an audit was conducted of screening results the last year before CAD was introduced and 2 years after. There was a 13% increase in the cancer detection rate with CAD, while the callback rate rose only slightly, from 8% to 8.4%. As in the Plano study, most of the cancers were detected in stage 0 or I.

Tommy E. Cupples, MD, medical director of mammography services, says the Breast Center uses CAD as a second reader. He thinks this practice will become routine.

“The hurdle to the adoption of CAD was that until last year, it was not reimbursable. It was hard to justify the additional cost of CAD for mammography, which is under-reimbursed in any case,” he notes. “However, now that Medicare has assigned both technical and professional reimbursement codes, there will be, if not a large incentive, at least not a large disincentive, for people to use CAD.”

Lung Cancer Screening?

Lung cancer kills more people (and more women) in the United States every year than does any other cancer, and we can easily identify the persons at high risk. So why is lung cancer screening not routine?

Figure 1. A digital chest radiograph with CAD results indicating two suspicious findings: the arrow over the right upper lung correctly indicates a small cancer. The arrow on the left lower lung indicates a normal blood vessel, which is a typical false-positive result. Even this imperfect level of accuracy, CAD has been shown to improve radiologists’ diagnostic accuracy. Images courtesy of Heber MacMahon, MD, University of Chicago.

Earlier trials such as the Mayo Lung Project showed no reduction in mortality in smokers who were screened even though more cancers were found and survival was better. The authors of a recent update of the project4 argued that the increase in survival meant that screening was identifying lesions of “limited clinical relevance.”? Today, that interpretation is widely rejected. Moreover, critics point out that chest radiography, which was used in all of these studies, has low sensitivity for small pulmonary lesions, so that the earlier results have little relevance in the era of low-dose spiral CT.

An argument that is more difficult to refute is that lung cancers metastasize so early that even small lesions probably are high stage. However, there is evidence of benefit from detection of small cancers.? The 5-year survival rate of patients with stage I non-small-cell lung cancer is 60% to 80% after radical resection, whereas similar patients who do not have surgery have a 5-year survival rate of only 10%.5 Moreover, the National Cancer Institute New York study, which enrolled 10,000 smokers who received chest radiographs with or without sputum cytology, found a 5-year survival rate of 35% among the patients who were found to have lung cancer during the trial.? The corresponding rate for the United States during the period of the study was 13%.6 As Dominioni and colleagues (and others) have written, the present policy in North America and Europe of discouraging lung cancer screening “is in urgent need of reconsideration.”5 The US National Cancer Institute hopes to show the value (or otherwise) of screening in the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening trial, but as chest radiography without sputum cytology examination (to detect airway cancers) is the method being used, a finding that screening is of no value may well be rejected by many pulmonologists.

Computer-aided detection might contribute substantially to examination of chest images and could be crucial if screening ever becomes routine, given the enormous number of people in the target population.? Computer analysis of digital chest images is being explored, and applications have been filed with the FDA for digital radiography systems for the chest.

A team of investigators from the University of Chicago was the first to test CAD for analysis of digital and CT chest images (see figure 1).7 They recently showed that an artificial neural network may help in the distinction of malignant and benign pulmonary lesions. A series of chest radiographs showing 34 primary lung cancers and 22 benign lesions were digitized with a pixel size of 0.175 mm and a 10-bit gray scale and examined by an artificial neural network. The network was more accurate than the radiologists in assessing the nature of the lesions, as shown by receiver operating characteristic analysis.8

The University of Chicago team also is working on automated detection of lung nodules by CT. In August, it published a report on the performance of its method on 43 spiral CT scans, showing a sensitivity of 70% for the whole series and 89% for the 20 examinations having only one or two nodules. In the whole series, the system averaged 1.5 false-positive identifications per section.8,9

Heber MacMahon, MD, interim chair of the Radiology Department and chief of the Chest Section at Chicago, noted that routine use of CAD could make thoracic radiologists far more efficient.

“CT images of the chest are acquired at high resolution, which means there is a large number of images that must be reviewed looking for small abnormalities; ie, early lesions,” he explains. “This is tedious and inefficient if you do it the traditional way. The computer could help, provided there is a good balance between sensitivity and specificity. If a program indicates several suspect lesions on every section, it creates more work for the radiologist than it saves and will not be used.

“To some extent, the desired specificity and sensitivity are a function of the radiologist’s confidence in his or her judgment,” he continues. “Some readers will be more tolerant of false-positive results (high sensitivity), whereas others will prefer higher specificity.”

A screening method for lung cancer that does not examine the airways will overlook many cancers, which is why some trials have included sputum cytology in screening examinations. Here, too, the computer may help. Virtual bronchoscopy-the use of CT data to create images of the interior of the bronchial tree-has been growing in popularity, particularly with the newer scanners capable of submillimeter resolution. The technique has several advantages in addition to noninvasiveness. For example, it permits assessment of the lumen beyond a narrowed area and depicts the relations between the lesion in the airway and adjacent structures.10 In one trial in 20 normal subjects, virtual bronchoscopy using narrow beam collimation and maximum pitch depicted the lumen of the bronchial tree as far as sixth-order branches.11 Seemann and Claussen of the Eberhard-Karls University of Tubingen, Germany, believe that virtual bronchoscopy may substitute for fiber-optic bronchoscopy in some patients, particularly in follow-up after treatment.12 Those investigators also posit a role in guiding biopsy, surgery, and palliative therapy.

Amid this enthusiasm, there are cautionary voices. Neumann and colleagues noted that the quality of the examinations depends heavily on the CT protocol used and the experience of the radiologist with perspective volume rendering. Greater sophistication of software might alleviate some of these problems, but there are larger issues.

“The startling visual appeal of these 3-D portrayals of a patient’s airway and mediastinal anatomy and the prospects of exploring this information in real time do not establish its clinical role,” wrote Haponik and colleagues.13 “Whether this added perspective translates into tangible benefits for patients ? has yet to be proved.”

Other Potential Applications

Screening for cancer is not the only application for which CAD may be valuable. The University of Chicago team demonstrated that artificial neural networks have a sensitivity and specificity of approximately 90% in detecting interstitial lung disease on digitized chest radiographs.13 Virtual arthroscopy was demonstrated at last year’s meeting of the Radiological Society of North America (RSNA).14 (Virtual colonoscopy will be discussed in another article in this series.)

Pulmonary emboli are a common and potentially fatal complication of many conditions. The standard imaging study is the ventilation-perfusion (V/Q) lung scan, but the results are not always definitive. Investigators in the Division of Nuclear Medicine at Massachusetts General Hospital used artificial neural networks, comparing their interpretations with those of three experienced radiologists of scans from 100 patients with normal chest radiographs.15,16 The computers were as good as the radiologists, and “[h]uman observers can improve their interpretations by incorporating computer output to formulate diagnostic prediction,” the investigators concluded.

At the RSNA meeting last year, a team from Stanford University described the use of CAD with ultrasonography to identify and measure carotid artery stenosis.17 The three-dimensional images were correlated with the ECG to determine the phase of the cardiac cycle, and the cross-sectional areas of the carotid arteries were classified into four grades reflecting the degree of narrowing. The condition of the vessels in 17 of the 19 subjects was correctly identified. The investigators noted that this study, which is still experimental, required only 12 minutes, approximately a quarter of the time needed for a standard bilateral duplex ultrasound examination of the carotid arteries.

Computers will not replace radiologists, but they are helpful tools. As CAD is validated in more parts of the body, double reading computer screening of images may become routine even in practices with small staffs. Patients will be the beneficiaries.


Judith Gunn Bronson, MS, is a contributing writer for Decisions in Axis Imaging News.


  1. Roehrig J, Castellino RA. The promise of computer aided detection in digital mammography. Eur J Radiol. 1999;31:35?39.
  2. Birdwell RL, Ikeda DM, Shaughnessy KF, Sickles EA. Mammographic characteristics of 115 missed cancers later detected with screening mammography and the potential utility of computer-aided detection. Radiology. 2001;219:192?202.
  3. Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: prospective study of 12,860 patients in a community breast center. Radiology. 2001;220:781?786.
  4. Marcus PM, Bergstralh EJ, Fagerstrom RM, et al. Lung cancer mortality in the Mayo Lung Project: impact of extended follow-up. J Natl Cancer Inst. 2000;92: 1308?1316.
  5. Dominioni L, Strauss GM, Imperatori A, Rovera F, Dionigi G. Screening for lung cancer. Chest Surg Clin North Am. 2000;10:729?736.
  6. Melamed MR. Lung cancer screening results in the National Cancer Institute New York study. Cancer. 2000;89(suppl 11):2356?2362.
  7. MacMahon H, Engelmann R, Behlen FM, et al. Computer-aided diagnosis of pulmonary nodules: results of a large-scale observer test. Radiology. 1999;213:723?726.
  8. Nakamura K, Yoshida H, Engelmann R, et al. Computerized analysis of the likelihood of malignancy in solitary pulmonary nodules with use of artificial neural networks. Radiology. 2000;214:823?830.
  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. Aquino SL, Vining DJ. Virtual bronchoscopy. Clin Chest Med. 1999;20:725?730.
  11. Neumann K, Winterer J, Kimmig M, et al. Real-time interactive virtual endoscopy of the tracheo-bronchial system: influence of CT imaging protocols and observer ability. Eur J Radiol. 2000;33:50?54.
  12. Seemann MD, Claussen CD. Hybrid 3D visualization of the chest and virtual endoscopy of the tracheobronchial system: possibilities and limitations of clinical application. Lung Cancer. 2001;32:237?246.
  13. Haponik EF, Aquino SL, Vining DJ. Virtual bronchoscopy. Clin Chest Med. 1999;20:201?217.
  14. Ashizawa K, Ishida T, MacMahon H, Vyborny CJ, Katsuragawa S, Doi K. Artificial neural networks in chest radiography: application to the differential diagnosis of interstitial lung disease. Acad Radiol. 1999;6:2?9.
  15. Penrod BJ, Lane MJ, Starkey M, DeBerardino T. CT and MRI ?virtual arthroscopy?: early experiences and correlation with conventional arthroscopy [abstract 0901MK-e]. Radiology. 2000; 217(P):679.
  16. Scott JA, Palmer EL, Fischman AJ. How well can radiologists using neural network software diagnose pulmonary embolism? AJR Am J Roentgenol. 2000;175:399?405.
  17. Napel S, Xu H, Paik DS, Sumanaweera TS, Hossack J, Jeffrey RB Jr. Computer-aided detection of carotid disease using 3D ultrasound [abstract 586]. Radiology. 2000; 217(P):342.