NEITHER cause nor cure is known for breast cancer, a serious disease that may affect one out of every nine women in the United States. Early detection is the only known means for increasing a victim's chances for survival; mammography is currently the best means of cancer detection in women showing no symptoms.
The power of mammography is proven. Yet, some breast cancers are missed, usually because the cancer is not imaged or because its indications in the image are too subtle to be recognized. The difficulty of visually detecting the cancer's subtle warning signs (in particular, sorting out significant microcalcifications--the calcium-rich deposits that are clues to malignant breast cancers) point to the need to improve image quality and the means of interpreting mammograms.
In 1991, help for improved breast cancer detection came from an unexpected source--Lawrence Livermore scientists and engineers working on national defense projects. They began to recognize that their technologies had important medical applications. Clint Logan, an engineer with expertise in materials imaging, an important aspect of nondestructive evaluations (see article in this issue), proposed using digital computer analysis on film mammograms. His proposal was carried out in a three-part project, first described in Energy and Technology Review, Nov.-Dec. 1992, pp. 27-36. The first part was to digitize mammograms, that is, to convert the data on the film record into numbers, applying a high spatial and contrast resolution to the entire mammogram. When digitized, data could be displayed with a variety of contrast settings, which allow clearer viewing than film studied over a light box.
The second part of this work was to develop computer algorithms to automatically detect microcalcifications in the digitized mammograms. The objective was to provide a "mammographer's assistant" that would quickly and objectively detect and flag microcalcifications for radiologists and doctors. The algorithm, developed by biomedical image processing specialist Laura Mascio, first performs two types of high-frequency analysis on a digitized image. One procedure extracts contrast (intensity difference) information, saving structures that have abrupt changes in brightness (from edges, for example) and are larger than several pixels in size. The other procedure extracts spatial, or size, information and thus saves small, textured structures.
Adding together what has been preserved by the two high-frequency analyses produces an image that is brightest where it contains detail common to both. When a selective erosion or enhancement (SEE) filter is applied over this image, it further reinforces image pixels that show strong evidence of belonging to a microcalcification and erodes pixels that show otherwise. The method developed by Mascio forms the basis of a computer algorithm that distinguishes between microcalcifications and mimicking spots, such as specks and flecks on the film. It was the first microcalcification-detection algorithm to use a gray-scale morphology for extracting frequency and texture information. It served as a model for further development of mammography screening algorithms.
The third part of the project was the design of a filmless, directly digital mammography system. Such a system would provide information and detection superior to the conventional film-based system, yet it would require a lesser x-ray dose to the patient. In collaboration with Fischer Imaging Corp., Logan and Jose M. Hernandez, another Livermore engineer, developed a digital screening unit with a novel x-ray source that can be adjusted for each patient's body size and an image detector that uses a charge-coupled device camera. Early trials indicate that this system yields images with better signal-to-noise ratios than conventional x rays. And because the images are digital, they can be manipulated in terms of contrast, magnification, and area of interest for the best view.

Improving Detection Algorithms
Algorithms having better sensitivity lead to earlier diagnosis of breast cancer and improved long-term survival. Algorithms having improved specificity (that is, they can separate suspicious spots that turn out to be benign from those that are malignant) mean fewer unnecessary biopsies and thereby less cost and less patient anxiety. However, sensitivity must be retained when improving specificity; otherwise, early, curable cancers could be missed.
In recent years, several other institutions have developed algorithms for computer detection of breast cancer. Until recently, however, there has been no way to compare the different algorithms because each research group has tested its own algorithms on different sets of film images that have varying degrees of diagnostic difficulty. Comparison of their relative performance is important because, in many cases, only partial records have been digitized.
To provide a standardized algorithm evaluation tool, Mascio and other Lawrence Livermore scientists began collaborating in 1995 with researchers from the University of California at San Francisco (UCSF) to compile a library of mammograms that could be used to test detection sensitivities and specificities. They used UCSF screening data of patients whose identities had been obscured. A total of 50 patient cases were selected to represent different categories: 5 normal, average, healthy cases; 5 normal but difficult cases (e.g., with implants, asymmetric tissue); 20 cases with obviously benign microcalcifications; 12 cases of suspicious but benign microcalcifications; and 8 cases of a biopsy-proven, malignant cluster of microcalcification. A radiologist then worked with the Livermore team, using all available clinical information, to annotate the mammograms.
The library is a first step toward a meaningful comparison of microcalcification-detection algorithms. The completely digitized mammograms have been put onto a CD/ROM in binary data format (see photos below) to make them available for other researchers. Images will be available soon on Lawrence Livermore's Web site (http://www.llnl.gov/).
Another problem with digital mammography is that its very large data files can present storage and processing problems, especially for small clinics with limited computer resources. Digital mammography usually records four views per patient, each taking up 200 megabytes of computer memory. To make this technology more efficient and practical, Mascio has proposed a way to compress mammogram files by factors of 10 to 30 without sacrificing image detail or diagnostic accuracy. Furthermore, it requires no decompression time when data are retrieved for viewing or analysis.
Generally, the more data are compressed, the more the data values differ from their original form once they are decompressed. Mascio's approach, called dynamically lossless compression, avoids wholesale data compression and instead selectively assigns the most data space (i.e., provides the highest spatial resolution) to the features that must be depicted in the most detail, such as detected microcalcifications. Less important features--such as background, healthy, nonglandular tissue--are given coarser resolutions. Thus, an image may contain many different spatial resolutions, each appropriate to the significance of the particular feature, and all based on mammogram-specific knowledge. This compression approach parallels human visual inspections of mammogram film--radiologists use a magnifying glass to get a higher-resolution view for studying microcalcifications, but they inspect larger abnormalities without the magnifier and by standing at a distance from the mammogram.






For the Next-Generation System
As a result of this collaboration, four direct-digital screening systems produced by Fischer Imaging Corp. have been installed at sites around the U.S. Even as they are being introduced to the general population, Jeff Kallman, a Lawrence Livermore engineer, is starting research on the sensors for a new generation of mammography screening. He proposes to generate three-dimensional images of soft breast tissue speedily and painlessly with linear ultrasonic diffraction tomography. Because breast tissue has neither large sonic variations nor appreciable multiple scattering, linear imaging techniques can be used. There is some evidence that cancerous tissue has sound speed and attenuation properties different from normal tissue; the hope is that such an imaging system will be able to distinguish between them.
Data collection would be done while the breast is immersed in water or gel, bypassing the breast compression that makes conventional mammography uncomfortable and even painful for some women. Furthermore, it would involve no ionizing radiation, thus eliminating concerns about x-ray exposure. With appropriate data-acquisition technology, which Kallman is investigating, breast cancer screening in the future would be done quickly as well as safely.
-- Gloria Wilt

Key Words: breast cancer, data compression, detection algorithms, digital mammography, linear ultrasonic diffraction tomography, mammogram library, microcalcifications.

For further information contact J. Patrick Fitch (510) 422-3276 (jpfitch@llnl.gov).


Back to December 1997