Automatic assessment of scintmammographic images using a novelty filter.

99mTc-sestamibi scintmammograms provide a powerful non-invasive means for detecting breast cancer at early stages. This paper describes an automatic method for detecting breast tumors in such mammograms. The proposed method not only detects tumors but also classifies non-tumor images as "normal" or "diffuse increased uptake" mammograms. The detection method makes use of Kohonen's "novelty filter". In this technique an orthogonal vector basis is created from a normal set of images. Test images presented to the detection method are described as a linear combination of the images in the vector basis. Assuming that the image basis is representative of normal patterns, then it can be expected that there should be no major differences between a normal test image and its corresponding linear combination image. However, if the test image presents an abnormal pattern, then it is expected that the "abnormalities" will show as the difference between the original test image and the image built from the vector basis. In other words, the existing abnormality cannot be explained by the set of normal images and comes up as a "novelty." An important part of the proposed method are the steps taken for standardizing images before they can be used as part of the vector basis. Standardization is the keystone to the success of the proposed method, as the novelty filter is very sensitive to changes in shape and alignment.