Technology of computer-aided detection

When using CAD with film-screen mammography, the images are acquired in the standard manner. Once finished, the films are fed into the processing unit to be digitized and processed by the algorithms. The digitized films are then sent to the viewer. The radiologist then reads the mammogram without any CAD markings and makes an interpretation. Once the reading is complete, the radiologist presses the button to activate the CAD display. This shows the search results on a monitor without marking the original film. The radiologist then reviews the ROIs indicated on the CAD review and updates the interpretation if necessary. Alternatively, the system can print out a hard-copy version of its findings, and the radiologist can refer to that rather than viewing the CAD results on screen.

With digital mammography, there are no hard-copy films to be digitized, so the initial soft-copy review is conducted as usual and then the CAD review is initiated. In all cases, however, it is essential that the radiologist fully read and interpret the findings before activating the CAD option or referring to the CAD findings.

The use of CAD in film-screen mammography has been shown to add 17 seconds of time for the radiologist and 1 minute, 21 seconds for the technologist. The added radiologist time is similar to the 16 seconds typically required to examine the mammogram with a magnifying glass.³

“We do not read mammograms online,” said Linver. “We batch-read our screening mammograms. We have the assistant hang the films and write down the result. This allows the reading radiologist to concentrate fully on the films, which increases accuracy and efficiency. We use the paper printout because by doing so we can read on all the multiviewers and are not tied to the multiviewer that has the CAD attached to it. The paper printouts then become part of the patient record,” he continued. “We read at the multiviewers, make up our minds, then look at the paper, and review the marks made by the CAD. Then we decide whether or not to change our initial im-pression. If we do change it, we do so right at that time. We always make our de-cisions first before we look at the CAD. I think if you try to do it any other way, your thinking is going to be overly colored, and you may overlook things that are not marked by the CAD.”

A 2001 study of more than 12,800 screening mammography patients compared reading results with and without the use of CAD.⁴ The study found that the use of CAD increased the recall rate from 6.5% to 7.7% but did not change the positive predictive value for biopsy (38%). Overall, there was a 19.5% increase in the number of cancers detected and the proportion of early-stage malignancy detection (defined as stages 0 and I) rose from 73% to 78%. The authors concluded that “the use of CAD in the interpretation of screening mammograms can increase the detection of early-stage malignancies without undue effect on the recall rate or positive predictive value for biopsy.”⁴ This figure of a 20% increase in detected cancers is now widely accepted.

Interestingly, this increased rate of detection is higher than that found with double reading, which has been shown to increase sensitivity by approximately 10% when compared with single reading.⁵

“The sensitivities for CAD’s ability to find early breast cancers have increased dramatically,” said Linver. “Its overall sensitivity is roughly 90% now, with highest sensitivities being for calcifications. It almost never misses calcifications on the film, but it still has some way to go in finding small, early invasive cancers—those that either consist of a mass or an area of architectural distortion. It finds appproximately 84% to 86% of known masses when they are presented to the computer. Overall, it is doing remarkably well as an adjunctive tool to the radiologist.”

“I think it has the most trouble finding architectural distortions without a true mass,” Linver continued. “But then again, so do we as radiologists. CAD is very much a reflection of what we do. In fact, that is how CAD was taught to find what it finds: by feeding it images with specific objects that we as radiologists believed were abnormal. So that’s how it learns: from us. Therefore, the things that are most difficult for us to find are going to be the most difficult for CAD to find as well.”

Another advantage of CAD is that it is always available and always at peak performance, unlike human readers. “The nice thing about CAD is that it is available and ready for action 24/7, whereas radiologists are subject to the frailties of the human condition,” said Linver. “We are sharp sometimes and we’re not so good at other times. The machine is going to be rock solid anytime it’s used. So in that sense, it has the potential to be much better than we are.”

As with many new technologies, the greatest barrier to widespread utilization lies in its cost. A CAD unit, depending on the model, can cost upwards of $200,000. Practices that perform a small volume of mammograms—and most mammograms are performed at facilities that do only a small volume of such studies each day—often will not be able to amortize that cost, even with the reimbursement that is currently available.

In addition, there is a learning curve with the use of this technology. “At first we are inclined to say, ‘If the computer finds it, then it must be significant,’” explained Linver. “But you have to remember that the computer is looking at these images in a vacuum and we are not. When we look at a mammogram, we have the advantage of being able to look at both views of each breast and make a decision based on both views. The computer has to make its decision on the basis of one view at a time. Not only that, it does not have the advantage of being able to review all the previous mammograms. So again, we have a tremendous advantage over the computer in our ability to sort out whether or not a marking is real, imaginary, or of no significance.” For this reason, the FDA-approved indications clearly state that the radiologist must interpret all images before reviewing the CAD data and the CAD information should be used only as an adjunct once the radiologist has formed his or her own clinical opinion.

Designed to assist the radiologist in detecting lung nodules, the system’s proprietary algorithms use automatic measurement and characterization information to examine the study in three dimensions and mark potential ROIs for additional review by the radiologist (Figure 4).

Two additional software packages were also recently approved for use with this system. One provides automatic three-dimensional registration and the ability to track lung nodule progression or regression over time. The other was designed to help radiologists see and evaluate pulmonary emoboli and other filling defects in pulmonary arteries.

Dr. Delgado has been using the CAD CT system for lung screening as part of the beta testing process. “The CAD primarily locates solid nodules of a small-to-medium range (anywhere from 4 to 30 mm), nodules that are of approximately fluid or soft-tissue density, those with calcium density, and nodules that are spherical or ovoid, or some derivation thereof,” continued Delgado. “The algorithm is not currently de-signed to pick up ‘groundglass’ opacities, just solid nodules.”

“This technology is particularly useful in patients who are being worked up for a suspicious nodule,” continued Delgado, “and in those being followed up from a chest X-ray. Also, in patients with a known and sustained nodule that is being monitored—which is currently accepted for certain nodules due to their size, morphology, or any associated history or findings in the patient—CAD can be used to look for potential changes over time. I think that is where there is a lot of use with CAD—that it can give you some information that the radiologist’s eye may have more difficulty in seeing, such as the volume, density, and radius of the nodule. Just the fact that there are nodules in the lungs may not truly be all the information available. We want to know more about each individual nodule and, specifically, how it changes related to therapy, such as chemotherapy, radiation, and other treatments.”

As with CAD for mammography, there is a learning curve associated with CAD for CT lung scanning. “Radiologists’ acceptance of CAD, I think, hinges on their ability to change, just as they did with positron-emission tomography (PET) or MR imaging or any new modality,” explained Delgado. “CAD is basically just an extension of what we have already been doing. Radiologists need to take a small leap-of-faith to delve into the new world of CAD—a bunch of digital information from a software program—and somehow incorporate that into their clinical practices. With CT systems going from 4-slice to 8-slice to 16-slice to 64-slice, radiologists do not have time to scrutinize every single image to such extreme detail. They need to still look at the case diligently, but they need to accept that they might need a little bit of assistance, that they might overlook certain things, such as potential early lung cancers.”

As in the use of CAD with mammography, it is essential that radiologists not use CAD as a crutch. All images must be read and clinical opinions must be formed before the CAD data is considered. “Radiologists still need to be in control of their ultimate interpretations and their conclusions based on their own observations of the CAD information,” said Delgado. “I think the biggest caveat that people need to consider is that the CAD does not become this all-knowing computer brain that knows everything. CAD does have false-positives. There are certain clinical settings in which the radiologist may completely disagree with the CAD or override the CAD findings; and that’s perfectly legitimate. Radiologists should not become afraid or always follow CAD findings as if they were gospel.”