Optimal Feature Functions on Co-occurrence Matrix and Applications in Tumour Recognition

The level of quality in segmentation and classification tasks, operating on biomedical digital images, is strongly affected by the ability of the feature functions to map the input structures on “highly separated” data sets. This level of separation is strictly dependent by the “discriminatory power” of the chosen feature functions, meaning with this expression the property of the functions to assign some very different values to different types of image blocks. This paper describes a possible theoretical approach addressed to the improvement of this “discriminatory power” of textural features in the particular case of features extracted using a co-occurrence matrix method. After the necessary references to the basic concepts, a formal definition of the “Discrimination Enhancement Problem” is stated. A new theorem is presented, solving the problem of discrimination among three different classes. The chosen application is the problem of tumour recognition in radiographic images of the brain and the experimental results support the effectiveness of this approach.