ELSEVIER Pattern Recognition Letters 18 (1997) 249-258 Pattern Recognition Letters A new minimum variance region growing algorithm for image segmentation Chantal Revol, Michel Jourlin * Laboratoire d'Image, Signal et Acoustique (LISA), CNRS EP92, Ecole Sup~rieure de Chimie, Physique et Electronique (CPE), 31 Place Bellecour, 69288 Lyon Cedex 2, France Received 3 November1995; revised 24 January 1997 Abstract Region growing is a very useful technique for image segmentation. Its efficiency mainly depends on its aggregation criterion. In the present paper, a new algorithm is proposed with a homogeneity criterion based on an adequate tuning between spatial neighbourhood and histogram neighbourhood. It differs from other techniques by reconsidering the pixel (or voxel) assignments on each step by a process which minimizes variance through special dilations. Thus, the region created by an initial seed can be non-connected and possibly does not contain this seed. Examples are given in dental surgery for 2D X-Ray images (and their associated 3D block) and for 3D images acquired by the Morphometre, the new 3D scanner constructed by GEMSE (General Electric Medical Systems). © 1997 Published by Elsevier Science B.V. Keywords: Image segmentation;Region growing; Variance minimization;Homogeneity;Histogram;Morphologicaloperations 1. Introduction Image segmentation is an important research area in image processing (e.g. Alnuweiri and Prasana (1992), Haralick and Shapiro (1985), Zucker (1976), Pavlidis (1988)), especially in medical applications. Image segmentation techniques can be roughly di- vided into three categories: measurement space guided spatial clustering, split and merge growing schemes and region growing schemes. The focus of this paper is on this third category. A region growing begins with a seed location and attempts to merge neighbouring pixels into this * Corresponding author. growing region until no more pixels can be added to it (e.g. Sivewright and Elliot (1994), Sekiguchi and Sano (1994)). The goal is to obtain a final region which corresponds to a whole object or a meaningful part of one. Image characteristics are used to group adjacent pixels in order to form regions. The process of merging pixels or regions to produce larger re- gions is usually governed by a homogeneity crite- rion. Various homogeneity criteria have been investi- gated for region growing (e.g. Copty and Ranka (1994), Wu (1993)). In the following, we will use the term pixel even if it can be applied to voxels. One of the most current region growing methods is by single linkage. It considers pixels as vertices in a graph. Neighbour pixels having similar properties are joined by an 0167-8655/97/$17.00 © 1997 Published by Elsevier Science B.V. All rights reserved. PII S0167-8655(97)00012-3