Capturing Microcalcification Patterns in Dense Parenchyma with Wavelet-Based Eigenimages

A method is proposed based on the combination of wavelet analysis and principal component analysis (PCA). Microcalcification (MC) candidate regions are initially labeled using area and contrast criteria. Mallat’s redundant dyadic wavelet transform is used to analyze the frequency content of image patterns at horizontal and vertical directions. PCA is used to efficiently encode MC patterns in wavelet-decomposed images. Feature weights are computed from the projection of each candidate MC pattern at the wavelet-based principal components. To assess the effectiveness of the proposed method, the same analysis is carried out in original images. Candidate MC patterns are classified by means of Linear Discriminant Analysis (LDA). Free-response Receiver Operating Characteristic (FROC) curves are produced for identifying MC clusters. The highest performance is obtained when PCA is applied in wavelet decomposed images achieving 80% sensitivity at 0.5 false positives per image in a dataset with 50 subtle MC clusters in dense parenchyma.