Computational detection of asymmetries between mammograms

Author(s)

Ricardo José Ferrari

• Document type: Doctoral Thesis
• Press: São Carlos.
• Institution: Universidade de São Paulo (USP). Escola de Engenharia de São Carlos (EESC/SBD)
• Defense date: 2002-04-01

Examining board members

Annie France Frère Slaets; Luciano da Fontoura Costa; Sergio Shiguemi Furuie; Nelson Delfino d'Ávila Mascarenhas; Antonio Carlos Roque da Silva Filho

• Advisor: Annie France Frère Slaets

Abstract

In this work, techniques are proposed for the automatic segmentation of mammograms and detection of asymmetries between left and right mammograms. The segmentation is performed by using three computational techniques for the identification of three important anatomical regions of mammograms: the skin-air boundary, the pectoral muscle, and the fibro-glandular disc. The first method focuses on the identification of the skin-air boundary by using an active contour model algorithm specially tailored for this purpose. In this stage, the skin-air boundary is demarcated, all artefacts outside the breast region are eliminated, and the region of interest for detection of the pectoral muscle is defined. In the next stage, the edge of the pectoral muscle is determined by using a multiresolution technique based upon a Gabor wavelets representation. Finally, a density breast model based upon a Gaussian mixture model is proposed for the representation of four categories of different density tissues in the breast. The fibro-glandular disc is identified by thresholding the density categories of the model. The methods proposed were applied to 84 images of medio-lateral oblique mammograms from the Mini-MIAS (Mammographic Image Analysis Society, London, U.K.) database. The evaluation of the skin-air boundary and the pectoral muscle edge were performed based upon the percentage of false-positive (FP) and false-negative (FN) pixels determined by comparison between the true contours and the contours automatically identified. The FP and FN average rates for the skin-air boundary and the pectoral muscle edge were, respectively, 0.41% and 0.58%, and 1.78% and 5.77%. Two radiologists subjectively rated the segmentation of the fibro-glandular disc and the results indicate that in more than 80% of the cases, the segmentation was considered acceptable for a Computer Aided Diagnosis purposes. Detection of asymmetries (continua) (continuação) is performed by using directional information, obtained from a multiresolution Gabor wavelets representation, and shape and density information, extracted from the fibro-glandular discs of left and right mammograms. In the directional procedure, a particular wavelet scheme with 2-D Gabor filters as elementary functions with varying tuning frequency and orientation, specifically designed in order to reduce the redundancy in the wavelet-based representation, is applied to the given image. The filter responses for different scales and orientation are analyzed by using the Karhunen-Loève (KL) transform and Otsu\'s method of thresholding. The KL transform is applied to select the principal components of the filter responses, preserving only the most relevant directional elements appearing at all scales. The selected principal components are thresholded by using Otsu\'s method and used to obtain the magnitude and phase of the image directional components. Rose diagrams computed from the phase images and statistical measures computed thereof are used for quantitative and qualitative analysis of the oriented patterns. A total of 11 features are also extracted from the segmented fibro-glandular discs of left-right mammograms, and the difference of each feature pair is used as a measure for detecting asymmetries. A total of 88 images from 22 normal cases, 14 asymmetric cases, and 8 architectural distortion cases from the Mini-MIAS database were used to evaluate the scheme. An exhaustive combination of the features along with the principal components analysis was used to select the best feature set. The classification was performed by using two Bayes\' classifiers (linear and quadratic) and the leave-one-out methodology. Average classification accuracy up to 84.44% was achieved. (AU)