Measurement of 3D structures such as tumor volume or measurement of structures for surgical planning is also becoming more common, although it lags behind 3D renderings. If such measurements could be generated more reliably and efficiently, they would likely become as common as 3D renderings.

Virtual endoscopy, including colonography, bronchoscopy, and angioscopy, has become a popular form of 3D imaging that appears useful for diagnosis and therapy planning. Colonography, perhaps the most advanced of this group, is primarily focused on detecting polyps, small cancerous or precancerous tumors arising in the wall of the colon. This is difficult because of bowel contents that are not cleared during the cleansing process and normal folds of the colon wall that can simulate polyps. Published results show sensitivities and specificities of 60% to 90%, depending on the patient population and size exclusion criteria.

It is possible, however, to align radiographs obtained months or years apart to subtract information and then detect subtle abnormalities.3,4 In some cases, the results can be impressive. But with current technology, too many cases fail to be useful in clinical practice. The superimposition problem can make correct alignment and subtraction of radiographs nearly impossible.

It is more feasible to align 3D images, particularly of rigid structures such as the head or bones, and perhaps even semirigid structures such as the torso. If this can be done, temporal subtraction may prove useful. A modified form of temporal subtraction that better accounts for partial volume effects, known as the boundary shift integral method,5 has also been used to detect subtle atrophy in MR scans of patients with certain types of dementia.

'3+'D DATA SETS

As noted, imaging devices are producing more and richer data sets that outstrip human resources to interpret. But I believe that these large data sets present an opportunity that can be harvested most effectively using computing power that is increasing at an even greater rate. Much of radiology training involves learning "signatures," patterns of intensity and shape that indicate certain disease processes. Traditional CAD has focused on these patterns from a single parameter-x-ray absorption.

But other imaging devices, particularly MRI and PET/CT, can provide multiple views of a disease or structure with different contrast properties. Precise anatomic alignment can combine images from different imaging devices and from prior examinations. Although this richer data set can be difficult for a human to integrate, it can be very useful for computer algorithms, particularly if knowledge bases about anatomy and pathology can be incorporated into them.

A simple, familiar example is a brain MRI in a patient with acute stroke symptoms. The probability of an acute stroke is very high if the examiner sees restricted diffusion, T2 intensity that is elevated, T1 intensity that is low to isointense, and restriction of these signal abnormalities to a portion of the cortex. If an examination from two months prior is available for this patient, and it shows no signal abnormality in this area, the probability that this is an acute stroke becomes even higher. On the other hand, if similar signal abnormality appeared on all sequences of the prior examination, other diagnoses should be considered.

There are three basic approaches to segmenting medical images. Thresholding is a familiar method in which voxels of a certain intensity range are assigned to certain structures. In and of itself, this is rarely useful, but adding spatial connectivity criteria can increase the usefulness substantially. Because thresholding and connectivity are fast to compute, interactive tools are feasible.

Region growing is similar to thresholding, but it adaptively changes the threshold based on local voxel values and requires connection with a seed point. If a gradual shift occurred in voxel values of a structure due to field heterogeneity in an MR scan, for example, absolute thresholds would not consistently identify borders of the structure. But adaptive methods would adjust the threshold as the region statistics slowly changed. These methods are sometimes employed in modern CAD but are becoming less popular.

A T1-weighted image, for instance, shows fairly good separation of gray matter, white matter, and cerebrospinal fluid but poor differentiation of tumors or some types of vascular disease. T2-weighted images have bright CSF and pathology, which can have similar intensities. Combining these two types of images permits separation of each tissue. The accompanying figure shows a scatterplot of intensity on T1- versus T2-weighted image types. Note that no perfect separation occurs between groups of points, likely as a result of partial volume effects, image shading due to imperfections in the field, and noise. Multispectral methods can fairly accurately classify tissues by identifying these clusters, however, even without clear separation.

There are many methods for determining where to separate clusters. Algorithms known as unsupervised classifiers determine the separations without any help from a human trainer. Supervised classifiers are given examples of each type of tissue and use those to determine the optimal boundaries.

Neural networks are often used for segmentation, including multispectral segmentation. Some approaches, such as the Kohonen self-organizing map,6 are unsupervised, determining how many clusters exist and where to separate them.

A human can also provide a training set of samples of tissues to more standard networks such as a multiplayer perceptron.7 The neural network uses these to help define clusters and optimizes the separation between classes by minimizing some error metric.

These techniques operate directly on the source images to separate the tissues. It is possible that better results can be obtained if intermediate images are created that maximally separate one tissue from all others. We have recently shown that this can significantly improve the classification of MR images compared with standard feature extraction methods.8

ANATOMIC ATLAS WARPING

A key component of making a diagnosis is knowing what to expect and looking for it. Traditional image processing methods and CAD have simply looked for signal patterns that denote disease. Any breast calcification of a certain shape and size would be considered worrisome for cancer, because the breast has relatively uniform texture. But most organs have a more complex architecture that confounds simple intensity-based methods.

Gray matter in the normal brain, for example, has somewhat higher signal than white matter on T2-weighted images and somewhat lower signal on T1-weighted images. Most white matter abnormalities have higher signal on T2 images than either gray or white matter and usually have lower T1 signal than white matter. Most white matter lesions thus appear to have gray matter at their margins. Indeed, intensity-based segmentation algorithms often label a ring of voxels around a white matter lesion as gray matter, even though humans know this is not anatomically normal.

Intelligence built into the algorithm indicating this anatomic information could help avoid this mistake. Humans can clearly delineate white and gray matter more precisely than do probability maps based on hundreds of individuals. Humans can also "interpolate" models of where structures should be even when no imaging features correspond. If a human knows, for example, that the sixth cranial nerve nucleus lies at a certain level in the pons and refreshes his or her memory using an anatomic atlas, the mental image is then warped onto a specific patient's anatomy to determine if a lesion does indeed affect a structure.

This process is becoming possible with computer algorithms. While 3D warping algorithms are not yet interactive in speed, they can often match an atlas with a high-resolution 3D image of the brain in a clinically reasonable amount of time. It then becomes feasible to routinely compute volumes of structures such as the frontal lobe, the left thalamus, and so on. This will lead to new diagnoses for diseases that have normal signal intensities on MRI.

A second approach uses all the voxels of an image. It uses a similarity term that pushes the deformation, an elastic force term that resists change in shape, and usually a term that resists sharp curvature. These parameters and others such as momentum and density allow the user to tailor the algorithm to a specific problem, but they may also limit the robustness. A universal set of parameters for medical images has not been found.

Perhaps more interesting is the possibility of detecting diseases that may not be perceptible with simple image viewing. Studies of populations may reveal that temporal lobe cortical thickness below a certain value indicates the presence of a certain type of dementia. An anatomic atlas that could map a specific patient onto a normalized atlas could first assist in creating standards and then assist in diagnosis that is not possible today.

DR. ERICKSON is an associate professor of radiology at the Mayo Clinic in Rochester, MN. This article is based on a presentation at the 2003 Symposium for Computer Applications in Radiology.

References

1. Lodwick G. A probabilistic approach to the diagnosis of bone tumors. Radiol Clin North Am 1965;3:487-497.

2. Lodwick GS, Turner AHJ, Lusted LB, Templeton AW. Computer-aided analysis of radiographic images. J Chronic Dis 1966;19:485-496.

3. Kano A, Doi K, MacMahon H, et al. Digital image subtraction of temporally sequential chest images for detection of interval change. Med Phys 1994;21:453-461.

4. Li Q, Katsuragawa S, Doi K. Improved contralateral subtraction images by use of elastic matching technique. Med Phys 2000;27:1934-1942.

5. Hajnal JV, Saeed N, Oatridge A, et al. Detection of subtle brain changes using subvoxel registration and subtraction of serial MR images. JCAT 1995;19:677-691.

6. Kohonen T. Self-organized formation of topologically correct feature maps. Biological Cybernetics 43:59-69, 1982

7. Minsky M, Papert S. Perceptrons: an introduction to computational geometry. Cambridge, MA: MIT Press, 1969.

8. Patriarche J, Manduca A, Erickson B. Improved classification accuracy by feature extraction using genetic algorithms Proc SPIE, San Diego, 2003.

9. Thirion J-P. Image matching as a diffusion process: an analogy with Maxwell's demons. Med Image Anal 1998;2:243-260.