Neural Networks and Prior Knowledge Help the Segmentation of Medical Images

Abstract

This paper describes some achievements in the segmentation of medical images using artificial neural networks. We have identified three main sources of a priori information available to help perform the task of medical image segmentation: anatomical knowledge about the imaged region, the physical principles of image generation and the "regulari ties" of biological structures. The exploitation of each of these forms of knowledge can be effectively achieved with suitable neural architectures, three of which are described in the paper. An important lesson learnt from using these architectures is that different kinds of knowledge unavoidably induce different limitations in the resulting segmentation systems, either in terms of generality or of performance. Our experience indicates that in several applications some of such limitations can be overcome through a careful exploitation and integration of available knowledge sources via proper neural modules.