Organ Segmentation by Means of Neural Networks

Abstract

In Medical Imaging, segmentation of organs and structures of interest is the basis for a large number of operations such as three-dimensional analysis, surgery planning and automated diagnosis. Unfortunately, image noise and the variability of biological structures contribute to make segmentation a very hard task. In order to achieve this goal, a proper fusion of information extracted from the image (bottom-up processing) with a priori knowledge of the scene in terms of models of interesting structures (top-down processing) is extremely important.

In this paper we propose a neural network based approach to the segmentation of structures of interest, which integrates bottom-up and top-down processing. The system includes four modules for each examined organ: a Focusing Module (FM) locates the organ of interest in the input image and the edges of the selected area are evaluated by a Low-Level Segmenting Module (LLSM). A High-Level Segmenting Module (HLSM) receives the output of the LLSM and gives a memorized generalized shape. Finally, an Organ Extracting Module (OEM) merges the high-level and the low-level data, producing the final organ segmentation. Both the FM and the OEM include feed-forward nets trained with the back-propagation algorithm, while the HLSM is an associative self-organized memory. The proposed approach has been applied to Computed Tomography and Magnetic Resonance images.