Catálogo de publicaciones - libros

A method of extracting salient image features in mammograms at multiple scales using the monogenic signal is presented. The derived local phase provides structure information (such as edge, ridge etc.) while the local amplitude encodes the local brightness and contrast information. Together with the simultaneously computed orientation, these three pieces of information can be used for mammogram segmentation including locating the inner breast edge which is important for quantitative breast density assessment. Due to the contrast invariant property of the local phase, the algorithm proves to be very reliable on an extensive datasets of images obtained from various sources and digitized by different scanners.

This paper describes an algorithm to segment mammo- graphic images into regions corresponding to different densities. The breast parenchymal segmentation uses information extracted for statistical texture based classification which is in turn incorporated in multi-vector Markov Random Fields. Such segmentation is key to developing quantitative mammographic analysis. The algorithm’s performance is evaluated quantitatively and qualitatively and the results show the feasibility of segmenting different mammographic densities.

Several studies have showed that increased mammographic density is an important risk factor for breast cancer. Dense tissue often appears as textured regions in mammograms, so density and texture estimation are inextricably linked. It has been demonstrated that texture classes can be learned, and that subsequently textures can be classified using the joint distribution of intensity values over extremely compact neighbourhoods. Motivated by the success of texture classification, we propose an fully automated scheme for mammogram texture classification and segmentation. The classification method first has a training step to model the joint distribution for each breast density class. Subsequently, a statistical comparison is used to determine the class label for new images. Inspired by the classification, we combine the so-called image patch method with a HMRF(Hidden Markov Random Field) to achieve mammogram segmentation.

Mammographic risk assessment is concerned with the probability of a woman developing breast cancer. Recently, it has been suggested that the density of linear structures is related to risk. For 321 images from the MIAS database, a measure of line strength was obtained for each pixel using the Line Operator method. The proportion of pixels with line strength above a threshold level was calculated for each image and the results categorised by Tabar pattern, Boyd SCC class and BIRADS class. The results indicated a significant difference between Boyd classes 1–3 (low risk) and classes 4–6 (high risk), and between most Tabar patterns and BIRADS classes.

Topological properties of the breast ductal network have shown the potential for classifying clinical breast images with and without radiological findings. In this paper, we review three methods for the description and classification of breast ductal topology. The methods are based on ramification matrices and symbolic representation via string encoding signatures. The performance of these methods has been compared using clinical x-ray and MR images of breast ductal networks. We observed the accuracy of the classification between the ductal trees segmented from the x-ray galactograms with radiological findings and normal cases in the range of 0.86-0.91%. The accuracy of the classification of the ductal trees segmented from the MR autogalactograms was observed in the range of 0.5-0.89%.

Two graph theoretic methods are used in conjunction with active contours to segment the pectoral muscle in 82 screening mammograms. To validate the method, the boundaries are also marked by four radiologists with different levels of experience in mammography. The simultaneous truth and performance level estimation (STAPLE) method is used to estimate the true boundary and to estimate the sensitivity and specificity of the segmentation schemes. The performance of one of the two algorithms is found not differ significantly from radiologists.