Contour-independent detection and classification of mammographic lesions. (March 2016)

• Record Type: Journal Article
• Authors: Casti, P.; Mencattini, A.; Salmeri, M.; Ancona, A.; Mangeri, F.; Pepe, M.L.; Rangayyan, R.M.
• Is Part Of: Biomedical signal processing and control. Volume 25(2016)
• Page Start: 165
• Page End: 177
• Publication Date: 2016-03
• Subjects: Breast cancer -- Computer-aided diagnosis -- Detection of masses and architectural distortion -- Gaussian curvature -- Gabor filters -- Contour-independent classification -- Spatial correlation -- 3D FROC; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2015.11.010
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: We present a multistage approach to detection and classification of mammographic lesions that is independent of accurate extraction of their contours. The ultimate goal is to discriminate malignant tumors from benign lesions and normal parenchymal tissue in a realistic scenario of lesion candidates automatically detected in mammograms. Local analysis of the Gaussian curvature and of the phase response of multidirectional Gabor filters is performed for identification of suspicious focal areas. The detection of lesions and the classification of malignant tumors are performed in series, respectively, via a differential approach to analysis of the tissue surrounding the candidates and via quantification of nonstationarity and spatial dependence of pixel values within circular and annular regions of interest. A unified 3D free-response receiver operating characteristic framework is applied for global analysis of the two binary categorization problems in series. The system was tested on a total of 2105 full-field digital and screen-film mammograms from three different datasets, including abnormal mammograms with 560 malignant tumors and 639 benign lesions, masses, or architectural distortion, and 1010 normal mammograms. For sensitivity of detection of malignant tumors in the range of 0.70–0.81, the range of falsely detected malignant tumors was 0.82–3.47 per image, with a series of two stages of classification, including stepwise logistic regression for selection of features, Fisher linear discriminant analysis, and two-fold cross-validation.