Veri Bilimi
Yazarlar: İdil ISIKLI ESENER, Şükriye KARA, Semih ERGİN
Konular:Mühendislik
Anahtar Kelimeler:Digital Mammography,Computer-Aided Diagnosis,Feature Extraction,Geometric Descriptor
Özet: Breast cancer is the leading cause of cancer-related deaths among women worldwide as well as being the most frequently diagnosed cancer type. Advances in technology introduce Computer-Aided Diagnosis (CAD) systems for breast cancer which gain importance in reduced mortality rate by the increased sensitivity of early diagnosis. Although any imaging technique can be adapted, mammography, with known effectiveness in early diagnosis, are mostly used in CAD systems for breast cancer. Hence, this paper focuses on design of a CAD system analyzing mammography images for breast cancer diagnosis and proposes a new approach for geometrical feature extraction for this system. The proposed scheme is verified on a subset of the publicly available Mammographic Image Analysis Society digital mammogram database. In the detection phase of this system, initially, adaptive median filtering is applied for noise reduction; artifact suppression and background removal is realized via morphological operations, and pectoral muscle removal is executed using a region growing algorithm. Then, Chan-Vese active contour modeling is utilized for the ROI detection. Thereupon, the center of gravity (CoG) of each ROI is determined, and a convex image is created by specifying 92 points, called as edge points, on the boundary curves of the related ROI. In the feature extraction stage of the diagnosis phase, the angles between each pair of edge points and the CoG, the Euclidean distance between edge points and the CoG, and the Euclidean distance between each pair of edge points are computed. These geometrical descriptors are utilized in the classification stage via the Random Forest classifier using the five-fold cross-validation technique. As a result, breast cancer diagnosis is achieved by an accuracy of 70.13%. Analyzing the overall confusion matrix constructed in the classification stage, it is clearly seen that although healthy and benign diagnoses are mixed, malignancy is diagnosed well by the proposed geometrical descriptors.