Inbreast-Database Masses Characterization



This work aims at investigating both shape and texture features in distinguishing malignant and benign breast tumours on mammogram images. Linear Discriminant analysis was applied as the classification methodology. Performance was assessed through accuracy, sensitivity, and specificity. The most relevant individual features were the normalized compactness and elliptic-normalized circumference. When combined, depth-to-width ratio together with either aspect ratio or elliptic-normalized circumference led to the highest performance. A sub-optimum algorithm to feature selection gave as best set of two features depth-to-width ratio combined either with aspect ratio or with elliptic-normalized circumference. Classification into benign versus malign by using any of these two subset of two features reached an accuracy of 89% with sensitivity of 100% and specificity of 79% .


Congresso Brasileiro de Engenharia Biomédica (CBEB) 2017

Cited by

No citations found