Home

Cover Story

Table of Contents

E-Newsletter

Article Archive

Editorial Calendar

Datebook

Writers' Guidelines

Orgs/Links

Opinion Polls

Reprints

Forum

For other articles and previous issues click here.

August 22, 2005

CT Lung Cancer Screening Building the Supporting Case
By J. K. Bucsko
Radiology Today
Vol. 6 No. 17 P. 20

The International Early Lung Cancer Action Program (I-ELCAP) looks at lung cancer screening differently than traditional randomized controlled studies—which, of course, are the gold standard of clinical trials.

In April’s issue of Radiology, Mayo Clinic researchers reinforced their earlier conclusions about lung cancer screening: positive LDCT exams don’t necessarily correlate with better outcomes. Too often, LDCT lung scans identify benign anomalies as suspicious growths, leading to unneeded expense, therapy, and surgery. And for the most aggressive cancers, in most cases, early detection, whether by LDCT or other means, simply makes no difference in either treatment planning or outcomes.

By repeating the same test year to year and tracking results, mammograms are an invaluable tool in helping physicians identify tumors in the early stages, when the choices for treatment are most numerous and effective. Mammography is a noninvasive, noncontrast, easily reproducible, and nowadays ubiquitous, exam covered by major insurers.

Unlike breast cancer, though, no mass screening technique has yet emerged for lung cancer. Now, recently released results of a long-term worldwide study seem to hold out promise for lung cancer screening—at least, according to the researchers who conducted that study. Others aren’t so sure, however. Earlier this year, both sides of the controversy resurfaced on the heels of new findings from the International Early Lung Cancer Action Program (I-ELCAP).

I-ELCAP Origins
Lung cancer, like breast cancer, is relatively common. It’s also typically asymptomatic until it reaches the most advanced stages. Roughly three-quarters of all lung cancers aren’t identified until they are in stages 3 or 4, at which point prognosis is poor: the five-year survival rates for patients diagnosed in those late stages are, respectively, 10% and 1%, according to I-ELCAP data. In fact, the five-year survival rate for all stages runs between 11% and 14%. On the other hand, when caught early in stage 1, the cure rate for lung cancer approaches 70%.

Through the 1970s into the mid-1980s, the Mayo Lung Project, under the auspices of the National Cancer Institute (NCI), conducted clinical trials of lung cancer screening techniques but determined they were insufficient for improving early diagnosis in the absence of other symptoms. That conclusion formed the foundation for all subsequent concepts for lung cancer screening trials.

Then in 1993, the initial ELCAP—first launched only in the United States—studied the efficacy of regular low-dose helical (spiral) computed tomography (LDCT) chest scans in identifying potentially cancerous lung lesions at the earliest stages. The expanded I-ELCAP program now comprises some 38 healthcare institutions worldwide, with major centers throughout the United States, Canada, Europe, and Asia, involving approximately 30,000 study participants.

Study Results
The original subjects were aged 60 or older and smoked cigarettes for at least 10 pack-years (one pack per day for at least 10 years, or two packs per day for five years). Participants had no previous history of cancer, were asymptomatic, and were able to undergo chest surgery (thoracotomy). Baseline CT exams involved a noncontrast scan from the thoracic inlet to adrenal glands with a single breath-hold up to 20 seconds at end inspiration after hyperventilation. Image slice thickness was 10 millimeters, with images reconstructed in overlapping 5-millimeter increments.

“To aid in diagnosing lung cancer by assessing tumor growth, follow-up diagnostic CT scanning was used and, if growth was identified, biopsy was done to confirm lung cancer. The [current] I-ELCAP protocol for the screening and follow-up tests is an updated version of the original ELCAP that reflects technologic and knowledge advances,” explains lead researcher Claudia Henschke, MD, PhD, of Cornell University Weill College of Medicine.

According to Henschke, successive yearly CT lung scans can detect a high percentage of cancers at stage 1, when they are most curable. In fact, she says, over the years I-ELCAP has found that deaths from stage 1 lung cancer “were surprisingly low after surgery, but only if treatment is pursued.” In patients who excised lesions subsequently found to be stage 1 cancers, the eight-year survival rate rose to 95%; in patients found to have a stage 1 lesion during a repeat scan, the survival rate jumped to 98%.

Those numbers, notes Henschke, are a significant improvement over the usual five-year survival rate for lung cancer. I-ELCAP researchers contend that annual screenings in smokers boost the cure rate to between 76% and 78%—compared with a cure rate of no more than 5% to 10% for smokers who don’t receive screenings.

Sharing Data
I-ELCAP uses many different scanner systems from a number of vendors, with baseline and follow-up CT results examined using various types of computer-aided diagnostic (CAD) programs. Participating sites use CAD systems from various vendors, including major CT makers such as Siemens and GE, as well as the independent developers described in the accompanying box.

Most importantly, says Henschke, all I-ELCAP sites are linked online via a Web-based interactive data management system originally developed by project researchers. The system uses a graphical user interface to guide data input through a series of “context-relevant” onscreen forms from initial contact through all subsequent encounters, including scheduling follow-up tests. The system prompts the user through full intake documentation, detailed medical history, reading CT images, and actions recommended and/or taken as a result of exam findings.

As staff at the clinical sites complete study data forms, that input is automatically transmitted to the site’s coordinating center, where the system checks for protocol conformity, completeness, and consistency. The central system includes service centers for terminology, emphysema index, public database, teaching files, and nodule growth and nodule detection images.

Using standard DICOM protocols, the system also supports full electronic transmission of CT scans and digital pathology “slide” images worldwide. So each I-ELCAP facility can immediately access data and images from every participating site around the globe. This allows for central reading, including automatic assessment of nodule volumes and growth rates. The computer system assigns centrally controlled, encrypted codes to ensure that individual patient identification is accessible only at the originating facility.

One key feature of the data management system is its capability for generating immediate reports summarizing each patient’s movement within the program. Collected reports are also used to develop teaching files illustrating exam findings made in the CT screening process, results of subsequent procedures, and final diagnoses.

At Issue
The idea of regular, preventive CT lung cancer screenings remains unendorsed by the medical establishment. In the April issue of the Radiological Society of North America’s journal Radiology, researchers from the Mayo Clinic reinforced their earlier conclusions about lung cancer screening generally, this time stating that positive LDCT exams don’t necessarily correlate with better outcomes. Too often, Mayo concludes, LDCT lung scans identify benign anomalies as suspicious growths, leading to unneeded expense, therapy, and surgery. For the most aggressive cancers, in most cases, early detection, whether by LDCT or other means, simply makes no difference in either treatment planning or outcomes, according to Mayo research.

Division still remains within the expert community—but, says Henschke, with data continuing to accumulate from numerous ongoing studies, some professional organizations are beginning to soften that position. Henschke says the U.S. Preventive Services Task Force has now reclassified the procedure, moving it from the previous “not recommended” status to the “individual decision” category. The American Cancer Society (ACS) also recently determined that CT scanning specifically for lung cancer detection should be a personal decision made by the patient with appropriate input from the physician, a recommendation with which the American Society of Clinical Oncology now concurs.

Still, formal guidelines from the NCI, American College of Chest Physicians, American Medical Association, and National Institutes of Health caution that available studies on spiral CT scanning for lung cancer cannot yet positively demonstrate a correlation between regular screening and reduced mortality.

Study Design Differences
Henschke believes much of the controversy around lung cancer screening success can be traced to the way the varying studies are set up. I-ELCAP, for example, is a nonrandomized study. All participants, known to be at high risk, receive both regular screening and prompt follow-up treatments (according to individual patient wishes). There is no control group of nontested/nontreated patients.

That methodology is in direct contrast to the traditional randomized study, which remains the gold standard for clinical trials. The NCI’s current National Lung Screening Trial (NLST), for example, is a fully randomized long-term trial involving roughly 50,000 subjects at 30 sites around the United States. It includes a control group of patients who will not receive CT scans, with the results of patients who do receive an annual CT lung screen every year for three years. The goal of the NLST is to determine whether CT scanning provides a 20% or greater difference in mortality rates from lung cancer, and, its researchers say, that kind of result can only be objectively measured against completely neutral control findings over time.

Conventional wisdom says that only randomized comparisons between screening-detected and traditionally-diagnosed cases to assess nodule growth and mortality rates ensures eliminating unintentional bias in favor of the screened group. In other words, without randomized blinding, a certain amount of selectivity can creep into the findings, specifically in terms of lead-time, duration, and overdiagnosis (Woloshin S, et al. Tobacco money up in smoke? Lancet. 2002;359(9323):2108-2111).

A New Approach
Henschke and her colleagues disagree with that view. Just as medical and information technology have advanced exponentially, so have the methodologies and understanding of statistics, she says. “We can now come up with more sophisticated study designs, but in screening the traditional approach has been the same for 30 years. And interestingly enough, it has resulted in nothing but controversy for those years.”

The problem with using “the same old study design” in the NLST lies, in part, with who analyzes the data, according to Henschke. Like the Mayo trials before it, the NLST focuses on the overall death rates from lung cancer within the entire study population. Henschke believes doing so misinterprets the total results because inevitably some first screenings will identify patients already in advanced late stage cancer. (For the complete argument, see Miettinen OS, et al. Mammographic screening: No reliable supporting evidence? Lancet. 2002;359:404-405; and Henschke CI, et al. Early Lung Cancer Action Project: Overall design and findings from baseline screening. Lancet. 1999;344(9173):99-105,86-87.)

She says, “The problem with the traditional randomized trial is the deaths that occur early in the trials are ones you can’t impact because the cancers have already spread by the time you find them. But suppose you find an early lung cancer [during the screening trial] and resect it; now you have to ask, ‘When would that person have died if you hadn’t found the cancer and removed it?’ This probably would have occurred as long as seven to 10 years later.

“The traditional approach is to count all deaths, but doing that dilutes your ability to find the difference that screening makes in the deaths that follow,” Henschke continues. Most importantly, she says, “the screening is not continued long enough to see its full benefit.”

Other Objections
In its most recent commentary, Mayo’s team cited a roughly 69% false-positive rate as a significant deterrent to recommending lung cancer screenings. False positives are uncalcified lung nodules that show up on screening but turn out to be harmless. The NCI’s Summary of Evidence for LDCT lung cancer screening notes that false-positive results in trials to date range from 20% to 50%.

Henschke concedes that CT “finds many nodules,” but emphasizes that “careful definition of those which require further work-up can limit this to below 15% for the first screening, and below 6% for subsequent screenings. Using follow-up CT with careful assessment of growth … the cancers can be identified within three months.” That protocol echoes recent thinking in other oncology research, including current NCI protocols.

The Need for Deeper Data
Henschke further argues that screening is an ongoing process requiring long-term follow-up. She believes that because I-ELCAP gathers and analyzes data following a single patient over eight to 10 years—and because the protocol is continually reviewed and refined—it provides a deeper database on disease progression than previous lung or other cancer studies.

Moreover, because all participating facilities have immediate access to each other’s input, the protocol is continually being reevaluated. The I-ELCAP committee meets twice annually and since its inception has refined the study protocols. “The more accurate we can be, the more we can cut down on the number of biopsies needed, way beyond the number being done in normal clinical practice right now,” Henschke says.

In the United States today, roughly 50% of all chest surgeries are for benign disease, she notes. “The point is to try to find those who have cancer when it’s still asymptomatic.” Regular LDCT scanning is arguably a less invasive, and potentially less costly, way to rule out malignancies. “We’re talking about finding lung cancer in people who may think they’re fine,” she says.

While readily recognizing that “mass screening can’t be a simple yes-or-no decision,” Henschke believes the only way to really cut through the controversy is to continue adding to the available data. “Let’s fund several studies, using different designs,” she says. “Each will provide some new information that will allow us to proceed in a careful, collaborative way.” She points out that both the NCI and the ACS endorse the need for alternative studies.

Expanding CT’s Future Use
So far all the major medical insurance carriers list CT scanning for lung cancer as “investigational” and thus generally not qualified for reimbursement. But as Henschke points out, “the truth is that many, many CT scans are done to screen high-risk patients for lung cancer every year, but they are listed as being for a cough or other definite symptom.” Growing understanding of that reality may possibly spur swifter acceptance of scheduled CT screening, at least for some patients.

Ultimately, says Henschke, “we should be finding cancers when they’re smaller, and so more treatable, [as well as] having to do less invasive procedures to [screen out] those patients who don’t have cancer. With CT progressing so rapidly, and our knowledge progressing, each year in essence you should be doing [lung cancer screening] better and better.”

— J. K. Bucsko is a freelance healthcare and technology writer based in Westville, N.J. She is a frequent contributor to Radiology Today.

Adding CAD
In any screening process, says International Early Lung Cancer Action Program lead researcher Claudia Henschke, MD, PhD, “the key thing is to try to determine nodule boundaries in a consistent manner at two points in time… Advances in CT with CAD [computer-aided detection] have made a dramatic impact on the ease of diagnosis.”

Although differing CAD programs use differing measurement techniques, their aim is the same: to improve and accelerate the process of either identifying nodules or quantifying changes over time. Basically, CAD programs function as a second reader for the radiologist and compare studies of the same patient taken on different dates. Using proprietary algorithms, the system locates the same nodule images on the prior and subsequent slides, then calculates and displays any changes, whether an increase or decrease in nodule size, density, or volume.

Even the most sophisticated CAD still may have difficulty distinguishing nodules attached to blood vessels from normal blood vessels—and of course, no CAD can distinguish benign nodules from malignancies. Still, today’s CAD benefits are already widely acknowledged. Studies have shown that it improves overall lesion detection, reduces reading errors and variations, increases reading throughput, boosts repeat screening efficiency, and improves remote reading capabilities. A retrospective analysis of lung CT/CAD results, including false-positive rates, is now underway at the University Health Network of Toronto that draws specifically on studies done using the systems described here.

R2 Technology, Inc.
The way a CAD algorithm calculates reference points in comparing nodules is crucial, because a patient is unlikely to be presented the same way twice, notes Terry Chang, director of CT marketing for California-based R2 Technology, Inc. Weight loss or gain, breath-hold technique, duration, position, scanner make and type, and many other factors affect image registration from one study to another, with the result that the reference points of the nodule in the body also shift.

A variety of ways exist for manually registering and comparing values, he explains. Generally these are 2-D–based, referencing (for example) the top, bottom, and bifurcation of the lungs, then making a linear extrapolation from those three points to find the same spot in the subsequent image study. In contrast, says Chang, R2’s ImageChecker CAD system collects a series of reference points in three dimensions. In his words, “Many people view CAD primarily as a tool to help detect round, spherical objects that are indicative of a nodule, but we help find, and automatically compare, actionable nodules.”

To accomplish this, the integrated CAD and AutoPoint temporal comparison feature for MDCT studies first segments out normal anatomy throughout the lung. “Our 3-D method of segmentation looks at vessels and airways, segments that out; looks at the chest wall, segments that out; and then what’s left should be the actual nodule,” says Chang. It then identifies reference points unique to each nodule: “Where it falls in the lung, how close it is to the heart, where it is next to the rib, how close to the apex of the lung.” The system then uses probability equations to determine whether a nodule is the same on both studies. Thus, Chang says, unlike with a linear system, the reference points can correlate any combination of images and slice numbers.

“When you’re measuring a nodule over time, you want to make sure you’re measuring that nodule consistently and in a standard way between readers,” Chang says. “So a lot of things must be calculated in temporal comparison—doubling time, volume change, density change, whether the nodule has actually decreased in size. [Making an accurate comparison] stems from reproducibility—on how well you measure it each time.”

With roughly 12 years of experience in medical CAD, R2 Technology is the oldest commercial player in the arena. Its ImageChecker CT Lung CAD (which can be integrated to PACS and in multifunction workstations) has already gone into a second version after receiving premarket approval from the FDA last year. According to R2 studies submitted to the FDA, the ImageChecker system reduced by 26% the number of solid lung nodules measuring 4 millimeters or more that had been missed on previous radiologist readings.

Medicsight
The Lung CAD from London-based Medicsight also segments the lung for imaging, using a “thresholding-based method,” according to the company’s director of radiology, Mary Roddie, MD, who is also consulting radiologist for London’s Charing Cross and Hammersmith Hospitals NHS Trust. The current Lung CAD package was previously called Lung CAR (computer-assisted reader); the name change, the company says, “will hopefully provide the backbone for clear and simple communication around what is becoming a very complex, yet exciting sector.”

Medicsight, she says, uses “a mathematical algorithm applied to the segmented lung in order to detect any spherical objects with diameter between 4 millimeters and 30 millimeters… To distinguish nodules from ‘non-nodule objects’ such as blood vessels, a sphericity filter is applied. This filter analyzes every voxel on the surface of each nodule to determine whether it and its neighbors form part of the surface of a sphere.” For each read, the radiologist can set the system to search out only the growths that fit a specific shape and/or density. That provides an extra level of control, especially, for example, when trying to identify nonsolid or low-density nodules.

Roddie explains, “As the filter value is reduced, voxels that form part of an increasingly less perfect sphere—a flattened oval, for example—are retained as prompts. Decreasing the numerical value of the filter thus enhances sensitivity for less spherical nodules” (although, she notes, “at the expense of more false-positive prompts”). Because Medicsight integrates CAD as a concurrent read, the radiologist encounters system prompts during the first interpretation of the images, “thus potentially improving sensitivity for nodule detection without prolonging reading time.”

In addition to speeding up the read overall, by targeting nodules for specific comparison, this capability helps increase physician confidence. As Roddie says, “[Radiology recognizes that] reader fatigue and large volumes of images contribute to perceptual error, [which] is more likely to occur in a screening setting where the prevalence of abnormality is low. [But] CAD software is not influenced by the prevalence of abnormality and does not suffer from fatigue.”

Medicsight, founded in 1999, offers its software through partnerships with PACS and multifunction workstation makers, notably Agfa and Vital Images. At present, Lung CAD installations are being clinically evaluated in six facilities in England, Canada, Spain, and the United States.

— JKB

Facing the Future of CT/CAD
Leon Rubinsztain, MD, radiologist at the VA Medical Center in Decatur, Ga., is poised to become one of the first users of R2 Technology’s ImageChecker Version 2.0, just as soon as the governmental paperwork is finalized. “We currently do second reads with another system, but it’s cumbersome and time-consuming,” he says. “With ImageChecker, the machine will detect nodules, automatically match them with the prior study, compare them, and give you a percentage increase or decrease in volume.”

Volume measurement is critical because nodules can appear to grow in only one direction; for instance, from superior to inferior while not growing anter-posteriorally. A human reader may dismiss a 1-millimeter change over six or 12 months as insignificant, for example, but the CAD software will register even that small growth as an increase in total nodule volume.

“The software is going to tell me that there’s a 20% increase in volume. That’s much more significant than saying the nodule has increased 1 millimeter in diameter, which a radiologist might bypass,” Rubinsztain explains. Unlike different doctors, with individual levels of training and expertise and each looking for different features, “the computer, working in tiny fractions and decimals, is always going to take the biggest number available in pixels.”

He expects the ImageChecker to improve departmental efficiency, as well: “If a patient has 10 nodules and I have to follow these 10 nodules every six months for two years, imagine how much manpower I have to put on comparing nodule per nodule from exam to exam. While having a software that immediately compares them, matches them, and calculates a percentage volume total change, taking all the nodules into account—not nodule for nodule, but the total for all 10 nodules, even if some increase while others decrease—significantly increases a doctor’s throughput.”

“We’re doing about 40 to 60 CTs a day now, and I read about 20 cases a day,” he says. “With CAD implementation, I anticipate increasing my reading by as many as four to five more cases a day. Our scanning time isn’t going to change, but if I boost my reading time, I can do more scans and other work.”

Rubinsztain believes the system’s automated report generation will further add to radiology productivity. “Being able to integrate the computer report with my dictated report will save a lot of time,” he says. “I don’t have to dictate every nodule, where it is, what’s happening to it, because [the details and measurements] are in the CAD report.”

More consistent measurements, more consistent comparisons, and faster growth detection will also make it easier to monitor a patient’s overall progress. “Oncologists want to know what’s happening as soon as possible,” Rubinsztain says. “Is this patient getting better with my treatment or not? With this technology, you can do that, and give a number to show [results].” He adds, “I think this software is going to become standard in the future, in every computer PACS or 3-D workstation.”

— JKB

Subscribe to Radiology Today Magazine!