Detection of masses in mammogram images using CNN, geostatistic functions and SVM

  • Authors:

  • Wener Borges Sampaio;Edgar Moraes Diniz;AristóFanes CorrêA Silva;Anselmo Cardoso De Paiva;Marcelo Gattass

  • Affiliations:

  • Federal University of Maranhão - UFMA, Applied Computing Group - NCA, Av. dos Portugueses, SN, Campus do Bacanga, Bacanga 65085-580, São Luís, MA, Brazil;Federal University of Maranhão - UFMA, Applied Computing Group - NCA, Av. dos Portugueses, SN, Campus do Bacanga, Bacanga 65085-580, São Luís, MA, Brazil;Federal University of Maranhão - UFMA, Applied Computing Group - NCA, Av. dos Portugueses, SN, Campus do Bacanga, Bacanga 65085-580, São Luís, MA, Brazil;Federal University of Maranhão - UFMA, Applied Computing Group - NCA, Av. dos Portugueses, SN, Campus do Bacanga, Bacanga 65085-580, São Luís, MA, Brazil;Pontifical Catholic University of Rio de Janeiro - PUC-Rio, R. São Vicente, 225 Gávea, 22453-900 Rio de Janeiro, RJ, Brazil

  • Venue:
  • Computers in Biology and Medicine
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

Breast cancer occurs with high frequency among the world's population and its effects impact the patients' perception of their own sexuality and their very personal image. This work presents a computational methodology that helps specialists detect breast masses in mammogram images. The first stage of the methodology aims to improve the mammogram image. This stage consists in removing objects outside the breast, reducing noise and highlighting the internal structures of the breast. Next, cellular neural networks are used to segment the regions that might contain masses. These regions have their shapes analyzed through shape descriptors (eccentricity, circularity, density, circular disproportion and circular density) and their textures analyzed through geostatistic functions (Ripley's K function and Moran's and Geary's indexes). Support vector machines are used to classify the candidate regions as masses or non-masses, with sensitivity of 80%, rates of 0.84 false positives per image and 0.2 false negatives per image, and an area under the ROC curve of 0.87.