Too much information?it may be the hallmark of this generation. We see it on TV, and we see it in business. Radiology has this problem, too. Ten years ago, radiologists dreamed about faster scanners. Now they’re here, and they have become a nightmare by producing images at a rate that is much faster than any human can integrate. And, with rapid adoption of 64-slice CT and parallel techniques in MRI, the growth trends are not slowing down?which is not a good thing when the industry is facing a shortage of radiologists.

Transforming the Radiological Interpretation Process (TRIP) is an initiative of the Society for Computer Applications in Radiology (SCAR of Great Falls, Va) that focuses on dealing with this problem.1 The TRIP committee agrees on several strategies to attack the problem of too much information (TMI), including image perception, image processing and CAD, navigation, and visualization. (All of these components of the TRIP initiative will be addressed at the SCAR 2005 Annual Meeting.)

It is well-known that the interpreting radiologist does not always see lesions that are present on images. Although it is probably not possible to reduce the error rate to zero, we could be increasing the miss rate by creating more images. It is arguable whether these “misses” are as significant when the slice thickness is small. However, few could argue that techniques to highlight findings, so that the radiologist would not miss them, would be preferable.

Understanding the perception process and how technology might improve detection is one obvious area of interest. Once perception is better understood, features that promote perception might be enhanced to reduce errors. Of course, one must be careful that false positives are not inordinately increased as a result. Therefore, methods more sophisticated than simple filtering will be essential. Indeed, early efforts at CAD were largely spatial filtering; today, CAD products incorporate sophisticated combinations of filters, neural networks, and the like.

One important question we must decide is whether good practice requires a human to observe every single slice of an examination. Clinical pathology already has implemented devices to address this issue; for some tasks, such as pap smear cytologies, machines are better at making imaging diagnoses. In fact, the best candidate tasks for CAD are those involving screening?where there are many examinations with well-controlled acquisitions and few positives. For the cases that are positive, it often might require a human to determine the correct diagnosis.

This model could be appropriate for radiology, such as with mammograms and screening thoracic imaging. This is the domain of traditional CAD algorithms that focus on detection, or what some are now calling CADe (for computer-aided detection). Other algorithms are being developed to assist in diagnosis?determining if a nodule is benign or malignant, or differentiating types of interstitial lung disease. This category has been termed CADx (for computer-aided diagnosis). For images other than mammograms, detection of cancer is a small part of the interpretation task. For these, CADe will do little to enhance radiologists’ productivity. Enter a third category of CAD called CAD-CD (for computer-aided change detection). If a prior examination is determined to be normal by a human, and if a computer can determine that a more recent image has no changes, or no changes on all but a small subset of images, it might be possible to significantly reduce the workload without degrading the quality of care.

When imaging some diseases, such as CTA or MRA of vascular disease, 3-D renderings can be very effective in reducing the number of images that must be reviewed. It is less clear if 3-D can replace 2-D for parenchymal imaging. Certainly, the industry has made forward strides?such as the replacement of surface rendering with volume rendering, and allowing interactive adjustment of the rendering and segmentation parameters, which could allow reliable perception of all relevant abnormalities?but it is still speculative. Until further advances are made, we must struggle forward with improved navigation through 2-D images. In some cases, viewing images in some plane other than axial could increase efficiency by increasing the coherence of structures?for example, viewing chest CTs in the coronal plane because the superior-inferior (SI) dimension of the chest is greater than the anterior-posterior (AP) dimension.

It also is becoming apparent to many that the mouse is not an adequate navigation device for medical imaging. Image interpretation requires adjusting the slice, plane, window width and level, the “focus” stack, and more. A joystick or mouse can provide a little more than two dimensions of input?the X and Y motion, plus 0 or more buttons, means that only one or two of the viewing parameters can be adjusted without switching modes. Of note, many computer games also have the need to adjust many parameters simultaneously, such as moving the player, manipulating objects, and adjusting viewpoints. Early efforts at using gaming input devices for driving the PACS display software have proved useful, and this appears to be a fruitful avenue of investigation. Other input devices include eye or hand tracking to allow a more natural input method, though the dimensionality problem remains with such options. Voice control is highly dimensional, but accuracy is suspect as the number of commands increases.

Using computer algorithms to reduce navigation demands is another perspective on the problem that is being researched. One common part of the navigation task is to align images from different series, possibly from different exams, for visual comparison and correlation. For 3-D (and 2-D multi-slice) data sets with reasonably thin slices, some computer algorithms can fairly routinely align one data set with another. The compute times are quite reasonable (less than 1 minute in most cases) and are usually quite accurate. The weaknesses are that the algorithms typically assume a fixed shape (the abdomen and intestines often break this rule) and are most effective with reasonable overlap (aligning the few overlap slices from a chest CT with an abdomen CT often will fail).

Don’t Miss It!

The Society for Computer Applications in Radiology (SCAR) is hosting its Annual Meeting on June 2?5 in Orlando, Fla. More than 120 exhibitors will showcase their products, technologies, and services. And the educational programming promises to explore the past, present, and future of medical imaging in a host of sessions. Medical Imaging will be there in full force; meet the staff at booth 1604. Visit www.scarnet.org for more information.

It is said that every cloud has a silver lining. It might be that the problem of TMI will force us to dramatically change the way we acquire, process, and interpret images. Computers have partly created the problem of too much data, because their exponential growth in power has made it possible to create large data sets. We must now creatively apply computer technology to help reduce these large data sets to manageable, understandable pieces of information. This can increase our productivity, which is crucial to the success of our profession.

Bradley J. Erickson, MD, PhD, is a neuroradiologist at the Mayo Clinic (Rochester, Minn), where, in addition to clinical duties, he directs the clinical imaging components of the electronic medical practice and is vice chair of Mayo’s division of information services. Erickson also is the chair of the SCAR 2005 Annual Meeting Program Committee.

Reference

1. SCAR TRIP Subcommittee: Andriole KP, Morin RL, Arenson RL, et al. Addressing the coming radiology crisis?the society for computer applications in radiology transforming the radiological interpretation process (TRIP) initiative. J Digit Imaging. 2004;17(4):235?243.