Detection of architectural distortion in mammograms using geometrical properties of thinned edge structures. Issue 1 (2nd January 2017)

• Record Type: Journal Article
• Authors: Lakshmanan, Rekha; T. P., Shiji; Jacob, Suma Mariam; Pratab, Thara; Thomas, Chinchu; Thomas, Vinu
• Is Part Of: Intelligent automation & soft computing. Volume 23:Issue 1(2017)
• Page Start: 183
• Page End: 197
• Publication Date: 2017-01-02
• Subjects: Architectural distortion -- breast cancer -- classification -- cuckoo search -- geometrical properties of edge -- structures -- mammogram -- pectoral muscle -- sensitivity -- specificity; Artificial intelligence -- Periodicals; Intelligent control systems -- Periodicals; 003.5
• DOI: 10.1080/10798587.2017.1257544
• Languages: English
• ISSNs: 1079-8587

Abstract

Abstract: The proposed method detects the most commonly missed breast cancer symptom, Architectural Distortion. The basis of the method lies in the analysis of geometrical properties of abnormal patterns that correspond to Architectural Distortion in mammograms. Pre-processing methods are employed for the elimination of Pectoral Muscle (PM) region from the mammogram and to localize possible centers of Architectural Distortion. Regions that are candidates to contain centroids of Architectural Distortion are identified using a modification of the isotropic SUSAN filter. Edge features are computed in these regions using Phase Congruency, which are thinned using Gradient Magnitude Maximization. From these thinned edges, relevant edge structures are retained based on three geometric properties namely eccentricity to retain near linear structures, perpendicular distance from each such structure to the centroid of the edges and quadrant support membership of these edge structures. Features for classification are generated from these three properties; a feed-forward neural network, trained using a combination of backpropagation and a metaheuristic algorithm based on Cuckoo search, is employed for classifying the suspicious regions identified by the modified filter for Architectural Distortion, as normal or malignant. Experimental analyses were carried out on mammograms obtained from the standard databases MIAS and DDSM as well as on images obtained from Lakeshore Hospital in Kochi, India. The classification step yielded a sensitivity of 89%, 89.8.7% and 97.6% and specificity of 90.9, 85 and 96.7% on 60 images from MIAS, 100 images from DDSM database and 100 images from Lakeshore Hospital respectively