Optimization of the dictionary size selection: An efficient combination of K-SVD and PCA to denoise and enhance digital mammography contrast

Abstract

Mammographic images are often characterized by a low contrast and a relatively high noise content, due to 3-D breast structures projection onto a 2-D image plane. These effects may hinder lesion detection. During the past decade, many techniques have been proposed to improve the mammography contrast. Nevertheless, some image regions might not be adequately enhanced, while others might be subjected to excessive enhancement. For that reason, we propose a method to denoise the images and enhance contrast uniformly. First, we used a machine learning method to create a sparse dictionary from the database, then we used the principal component analysis to reduce the size of the dictionary before decoding each patch of a given mammography. Finally, the algorithm was tested on MIAS and INbreast databases using the same parameters’ values for each image. The results show that the visibility of breast mass and anatomic detail were considerably improved compared to the wavelet method and the computation time is halved compared to the conventional sparse coding algorithm and the curvelet method.