Characterization of clustered microcalcifications using multiscale Hessian based feature extraction

Segmentation of microcalcifications (MCs) significantly influences the performance of shape-based method for the diagnosis of MCs, which continues to be a challenge as it tends to have high false positive results. Texture based characterization of MCs represents a possible alternative that does not require prior segmentation of MCs and may improve the positive predictive value of automated diagnosis of MCs. This paper presents a new approach to extracting textural features, specifically spectral measures, of mammographie MCs using multiscale Hessian filtering (or equivalently second derivative of Gaussian). Extracted features were individually ranked using Fisher-score criterion, which demonstrated the superior predictive ability of the normalized entropy. A set of mammographie regions (20 malignant and 13 benign cases) from the MIAS database were used to evaluate the classification performance of the proposed spectral features. Utilizing k-nearest neighbor classifier and ROC performance measure, the proposed Hessian based extracted features produced ROC curves with performance index Az = 0.83, which demonstrated the effectiveness of the proposed characterization scheme.