Kiefer-Panorama-Röntgen-Aufnahme (OPT) Auswertung mit Computer
Your bosom buddy
Aug 10th 2006
A new technology that may improve the detection of breast cancer
SUCCESSFUL treatment of breast cancer depends on early diagnosis. The most widely used test relies on a low dose of X-rays to generate detailed images of the organ. This technique, known as mammography, can show changes in the breast well before a woman or her doctor can feel them, and it has significantly reduced mortality from the disease.
Reading mammograms, however, is a tricky business. Some signs of cancer appear, to the eye, similar to normal tissues on a mammogram. By contrast, dense but healthy breast tissue can obscure tumours beneath it. As a consequence, radiologists relying only on their own judgment may fail to notice up to 30% of breast lesions during screening, even though two-thirds of those lesions are visible in retrospect.
Computer-aided detection (CAD) can help. It uses special algorithms to scan mammograms and alert radiologists to things that seem suspicious—a strategy known as feature extraction. CAD has substantially increased the number of tumours identified. It is, however, less than ideal. The sensitivity of the technique to potential abnormalities is often raised, in order not to miss anything important. But that introduces false positive results; in other words, some normal tissues are marked as suspicious. In addition, existing CAD systems do not provide any explanation about how they came to their conclusions. Without such an explanation, some radiologists are reluctant to accept a diagnosis at odds with what they think their eyes are telling them.
Georgia Tourassi, of Duke University Medical Centre, North Carolina, and her colleagues hope to overcome this reluctance to be overruled by a machine. They are developing a CAD system that not only detects cancer more accurately than existing ones, but also acts more like an intelligent colleague than a black box. Instead of relying on feature extraction, Dr Tourassi's technique works by comparing the images taken by a radiologist with a large collection of normal and abnormal mammograms held in a database. This still requires algorithms, but of a different type. She and her colleagues have developed template-matching algorithms to compare the intensity and distribution of pixels in different images. They have also created decision algorithms to determine, after it has been compared with the entire database, whether a region of interest on a mammogram is normal or cancerous.
The diagnosis of the cases in the database has been confirmed, either by biopsy or by long-term follow-up, so there is no doubt about their details. If a new mammogram is similar to known cases of breast cancer, this would give reason for suspicion. This is exactly how a radiologist relates a case to those he saw in the past.
Making a clean breast
Dr Tourassi has found that her system can reliably distinguish tumour masses from normal tissues, and has a lower rate of false positives than systems based on feature extraction. Also, crucially, it can explain to a radiologist how it reached its decision by showing him similar mammograms in the database. The radiologist is then in a better position to decide whether the computer's judgment is valid.
The knowledge-based system has another bonus, too. As mammograms of new cancer cases are added to the database it is looking at, it will become cleverer—just as radiologists and physicians become more experienced and skilful as they come across more patients. This is in contrast to feature-based CAD systems, which cannot adapt to new cases unless their algorithms are suitably modified. However, there is a potential problem in the long run. The knowledge-based CAD system has a huge demand for computing power and, as the database grows, it will get slower and less efficient.
Dr Tourassi, however, has been thinking about this problem. She suspects that by using only the most informative mammograms, it might be possible to keep the size of the database within reasonable limits. To decide which ones to select, she turned to a branch of science called information theory. This theory says that the amount of information in a system can be measured in terms of its entropy.
Entropy is actually a measure of disorder. In the context of image processing, it is an indication of the complexity of an image. For example, an image that is all black or all white has zero entropy; an image of a chessboard, which contains an equal number of regularly distributed light and dark pixels, has low entropy; images with more varied distributions of many intensity levels of pixel have high entropy and are considered more informative.
Entropy is the basis of a standard indexing strategy for images and is often calculated automatically when they are put into a database. It is therefore easy to rank a collection of images by their entropy. In a pilot study, Dr Tourassi and her colleagues tested whether a subset of high-entropy mammograms would work as effectively as using the whole database.
Their entire database contained 2,318 mammograms from the Digital Database for Screening Mammography collected at the University of South Florida, but they varied the size and content of the collection of images examined by their algorithms, by either randomly selecting a subset of mammograms or choosing those with high entropy. They measured the utility of the different sets of images by taking each mammogram within it in turn, and testing it against all the others.
Their conclusion was that testing against only the 600 most informative mammograms was as effective as using all 2,318 images. And the system took less than three seconds per query—far faster than a radiologist could manage. A mammogram-reading machine that not only mimics but surpasses human perception and can explain its diagnosis would truly be a girl's best friend.