L.A. da Silva, E. Del Moral Hernandez (Brazil), and R.M. Rangayyan (Canada)
Neural networks, Committee machine, breast cancer, breast masses, shape analysis, texture analysis, pattern classica tion
This paper addresses a new approach using a committee machine to classify masses found in mammograms as be nign or malignant. Three shape factors, three measures of edge sharpness, and fourteen texture features were used for the classication of 37 regions of interest (ROIs) related to benign masses and 20 ROIs of malignant tumors. The committee machine is a group of classiers used to resolve a difcult task. Committee members are typically neural networks. In this work, we used a group of multi-layer per ceptrons (MLPs) as a committee machine classier. The classication results were realized by combining the res ponses of these classiers. Experiments involving change in the learning algorithm of the committee machine also were conducted. The classication accuracy was evalu ated using the area Az under the receiver operating char acteristics (ROC) curve. The Az result for the committee machine was compared with the Az results obtained using MLP and single-layer perceptron (SLP) neural networks; in some cases, the result was also compared with the Az ob tained via linear discriminant analysis (LDA). In almost all cases, the committee machine outperformed the SLP, MLP, and LDA methods; for example, with the measure of acu tance, the Az values of the methods were, in order, 0.79, 0.53, 0.70, and 0.74.
Important Links:
Go Back
