Data Representation in a Neural Networks Application for Segmentation of Magnetic Resonance Imagenes

G. BERNASCONI; M. REYES; H. TORRES; B. DROZDOWICZ

Tandil, Argentina

Jornada; 29 Jornadas Argentinas de Informática e Investigación Operativa; 2000

Sociedad Argentina de Informática e Investigación Operativa

This paper presents a Neural Network (NN) application for registered human brain Magnetic Resonance Images (MRI) segmentation. The selection and preprocessing of input data set was analyzed in more details, taking into account that the quality of the network results depends directly on the characteristics of the training patterns. In this way data reliability was analyzed, considering the convenience of complementary software to improve the data acquisition process. The inclusion of context information is analyzed in order to improve the Neural Network results. This context information is obtained considering the values of all the neighbors pixels. In order to enhance the generalization capability of the resultant networks some noise injection into input patterns in back-propagation learning is analyzed. A Modular Neural Network architecture was adopted.