Mark Palacio is Executive Editor of Applied Radiology.
Since the inception of computer-aided detection (CAD) systems in mammography nearly a decade ago, the tools have been refined and better integrated into radiology workflow. Today, digital mammography systems have alleviated many of the workflow challenges posed by first-generation CAD systems. Bottlenecks typically occurred when film-screen mammograms had to be digitized and taken to a separate CAD workstation for processing. Now with an entirely digitized workflow, CAD results are available at the push of a button. CAD solutions have also expanded beyond breast imaging. The solutions are available for chest X-ray, chest computed tomography (CT), CT colonography, and magnetic resonance imaging (MRI) of the breast and prostate. With rapid advancements and refinements in CAD algorithms, are these tools ready for full-time clinical use?
In general, CAD is designed to help the radiologist focus on areas that may have been overlooked in the initial read. The challenge to a radiologist is to achieve a high sensitivity, meaning that the radiologist identifies every cancer on a mammogram, for instance. However, it is well-documented that radiologists cannot detect 100% of breast cancers on mammography. They routinely achieve 85% to 90% sensitivity so any tool that bolsters a radiologist’s sensitivity is worth careful examination.
“The main way CAD helps us is by reducing perception errors,” said Matthew Gromet, MD, JD, FACR, of Charlotte Radiology, Charlotte,N.C. “The CAD analysis highlights areas that may have been overlooked. It’s up to the radiologist to determine if those marks are clinically relevant.”
Perception is an incompletely understood science, according to Dr. Gromet who uses a CAD solution from Hologic (Bedford,MA). Consider a “Where’s Waldo” puzzle. Some individuals can look at the puzzle and instantly spot Waldo, while others labor over locating the striped-shirted wanderer. Similarly in medical imaging, there are some findings that are not initially perceived for numerous reasons.
The classic contention when validating the usefulness of mammography CAD had been that it raised nonexpert mammographers to the level of expert mammographers and that expert mammographers had a marginal performance increase. Dr. Gromet recently published a study that showed that even expert mammographers could significantly benefit from CAD. He examined performance of experts (those who read >200,000 mammograms per year). The study compared traditional double reading with single reading plus CAD.1
After traditional double reading, there was a recall rate of 11.9%, 88% sensitivity, and 3.7% positive predictive value, which was determined as a percentage of positive screening mammograms that resulted in a tissue diagnosis of cancer within 1 year (PPV1). Single reading with CAD yielded a 10.6% recall rate, 90.4% sensitivity and 3.9% PPV1. Single-reading with CAD contributed to a statistically significant lower recall rate compared with double reading (10.6% vs 11.9%, respectively).1
So CAD systems can benefit the interpretation of mammograms, and digital acquisition and reading environments can smooth the integration of CAD applications but these systems still alter the radiology workflow. The workflow challenge is presented when a radiologist has to review an image for a second time.
CAD systems must be evaluated on their ability to minimize false-positive marks because each false-positive must be reviewed and dismissed by the reader. In the realm of chest X-ray CAD, Heber MacMahon, MD, Professor of Radiology and Director of Thoracic Imaging, at the University of Chicago Medical Center, Chicago, IL, has been using a CAD solution from Riverain Medical (Miamisburg, OH) for chest X-rays and he has seen steady improvements in the amount of false-positive marks.
Dr. MacMahon and colleagues recently published a study in Radiology that retrospectively determined the sensitivity of and number of false-positive marks made with an early version of their Riverain CAD system. The system was tested on a series of lung cancers that were previously missed on chest radiography. The CAD system had an overall sensitivity of 35% (identifying 12 of 34 cancers) and it helped identify 7 of 23 very subtle (30%) and 5 of 11 (45%) relatively obvious radiologist-missed cancers. The CAD program marked an average of 5.9 false-positives per radiograph.2
The results are promising because the early version of the software used in the study commonly marked as many as 5 false-positives per radiograph. Mostly this is attributable to bone, which has been problematic for chest CAD applications. Newer software versions have reduced the number of false-positives to an average of <2 per study, according to Dr. MacMahon.
Chest CAD paradigm
While chest X-ray CAD can assist in detection of subtle lesions, its true benefit will be evident when it is applied to every chest X-ray performed.
“It is the only logical way to use this detection technology,” said Dr. MacMahon. “The greatest value is when we’re not looking for a lung lesion. It absolutely makes sense to run a routine CAD analysis on every single case.”
At the University of Munich Clinics, in Munich, Germany, radiologists run Siemens (Malvern, PA) chest X-ray and chest CT CAD on every procedure. According to Chest Radiology Fellow, Peter Herzog, MD, the extra step does not significantly slow down trauma workflow. The additional CAD results become an extra sentence in the radiology report and patients receive follow-up care for lung lesions after their emergent situations are resolved.
The CAD solutions have already been used in more than 100,000 imaging studies at the University of Munich Clinics. The system has detected missed lung nodules on numerous occasions, according to Dr. Herzog. Many times, the nodules were overlooked because of the emergent presentation and when imaging is not done with the intent to look for lung nodules radiologists can be easily distracted.
“The key is not to miss the lesion,” said Dr. Herzog. “As long as a computer helps you to find them it is ultimately on the radiologist to decide if the lesion is malignant.”
CAD in prostate and breast MRI
CAD has also become an invaluable aspect of breast and prostate MR procedures. The solutions allow radiologists to quickly synthesize information from large MRI datasets by applying software that calculates numerical values and renders a color-coded image overlay that depicts the biological processes taking place in malignant vs. benign tissue.
“CAD improves our efficiency,” said Russell N. Low, MD, Medical Director of Sharp and Children’s MRI Center, San Diego. “It doesn’t add to our work and it improves our accuracy and our confidence.”
After performing an MRI study, the CAD analysis on Dr. Low’s iCAD (Nashua, NH) software can be rendered on the CAD workstation or on the PACS workstation. Colorized CAD images are integrated into the initial radiological interpretation.
“When we do a perfusion study of a breast or prostate, we acquire hundreds of images,” said Dr. Low. “A prostate perfusion study is typically 368 images. For me to make sense of those hundreds of images is impossible. The CAD will process those images, evaluate the kinetics of how the tumor enhances and present that data in a quantitative and qualitative image—and, most importantly, the software presents the information in a way that is meaningful.”
Since the CAD results have become part of the patient report, Dr. Low’s referring physicians no longer distinguish between CAD-processed images and unprocessed images.
“I present these at tumor board every week and they don’t think of these as CAD-processed images; they think of them as simply a breast MR study,” said Dr. Low.
With CAD systems reaching new levels of technical sophistication and achieving better integration with PACS, the time is right to bolster departmental efficiency and patient care with these tools. From quickening radiology workflow by dramatically simplifying multiple-image datasets to providing a second pair of eyes when reviewing complex anatomy, CAD has demonstrated that it is making a positive impact in clinical use.
- Gromet M. Comparison of computer-aided detection to double reading of screening mammograms: Review of 231,221 mammograms. AJR Am J Roentgenol. 2008;190:854-859.
- Li F, Engelmann R, Metz CE, Doi K, MacMahon H. Lung cancers missed on chest radiographs: results obtained with a commercial computer-aided detection program.Radiology. 2008;246:273-280.