Classification Performance related to Intrinsic Dimensionality in Mammographic Image Analysis

In the problem of mammographic image classification one seeks to classify an image, based on certain aspects or features, into a risk assessment class. The use of breast tissue density features provide a good way of classifying mammographic images into BI-RADS risk assessment classes. However, this approach leads to a high-dimensional problem as many features are extracted from each image. These features may be an over representation of the data and it would be expected that the intrinsic dimensionality would be much lower. We aim to find how running a simple classifier in a reduced dimensional space, in particular the apparent intrinsic dimension, affects classification performance. We perform classification of the data using a simple k-nearest neighbor classifier with data pre-processed using two dimensionality reduction techniques, one linear and one non-linear. The optimum result occurs when using dimensionality reduction in the estimated intrinsic dimensionality. This not only shows that optimum performance occurs when classifying in the intrinsic-dimensional space but also that dimensionality reduction can improve the performance of a simple classifier.