Mammographic Mass Detection by Robust Learning Algorithms

Cao, Aize; Song, Qing; Yang, Xulei; Wang, Zhimin; Shui, Yan

doi:10.1117/3.651880.ch5

Book: Recent Advances in Breast Imaging, Mammography, and Computer-Aided Diagnosis of Breast Cancer

Editor(s): Jasjit Singh Suri, Rangaraj M. Rangayyan

Author(s): Aize Cao, Qing Song, Xulei Yang, Zhimin Wang, Yan Shui

Published: 2006

https://doi.org/10.1117/3.651880.ch5

Abstract

Cancerous tumors are hard to visually detect when they are embedded in or camouflaged by varying densities of parenchymal tissues, particularly dense parenchymal breast tissues. It is a challenging task to detect masses within such background tissues in screening mammography, which has been recommended as the most effective method for the early detection of breast cancer. In addition, masses with various sizes and shapes may fail to generate a template that presents the common geometric properties of tumors. A lot of research work based on various theories has been carried out to tackle the problems of computerized mass detection. Doi et al. have developed several methods for automatic detection of masses in mammograms. Other methods of computerized detection and analysis include the earlier work done by Li et al. They proposed a Markov random field approach that lies in the category of a region-based algorithm to do breast-mass detection. The algorithm was reported to have achieved 90% sensitivity with two false positives (FPs) per image. Petrick et al. proposed a two-stage adaptive density-weighted contrast enhancement (DWCE) filter in conjunction with a Laplacian-of-Gaussian edge detector for mass detection. They reported 96% detection accuracy at 4.5 FPs per image for 25 mammograms by using a set of morphological features. Texture features that are based on gray-level co-occurrence matrices were studied later for a data set of 168 cases. A detection accuracy of 80% was achieved at 2.3 FPs per image. Naga and Rangaraj proposed employing Gaussian smoothing and subsampling operations as preprocessing steps in mass detection. The mass portions are segmented by establishing intensity links from the central portions of the masses into the surrounding areas. A sensitivity of 81% was achieved at 2.2 FPs per image for mass versus normal tissue. Malignant mass versus benign case classification resulted in Az = 0.9 under the ROC curve for 26 masses. Gradient and texture analysis was used in Ref. 7 to classify masses into benign and malignant cases. Some methods were proposed for the detection of spiculated masses because of their high likelihood of malignancy.