Colloquium Biometricum

Vol: 41 Page: 73

Authors:

Title:
LARS regression in diagnosing mammograms: selection of variables for analysis

Language: English

Keywords:

fractal signatures

mammogram diagnostics

linear regression

reduction of predictors

forward search

least angle regression

lars

correct classification

cross-validation

Summary:
Diagnosing cancer in a mammogram is a difficult task. Our aim is to explore the usefulness of so called fractal signatures for this purpose. A fractal signature is given by a vector of p real numbers characterizing the roughness of a mammogram considered as a texture file. Fractal signa-tures of length 48 are considered. Are all of them relevant to make the 2–group diagnosis: non–cancer or cancer? To answer this question, we used the Least–Angle Regression (LARS) which is believed more stable than the traditional forward search. By 5–fold cross–validation we found that only a small subset of variables is relevant for the diagnosis. The considerations are illustrated using data from the MIAS data base.