Techniques for Medical Image Segmentation: Review of the Most Popular Approaches

Introduction

Digital image processing applied to the field of medicine offers numerous benefits. They include in particular improvement in the interpretation of examined data, full or nearly full automation of performed tasks, better precision and repeatability of obtained results and also possibility of exploring new imaging modalities, leading to new anatomical or functional insights. One of the most important steps involved in the process of image analysis is the segmentation procedure. This refers to partitioning an image into multiple regions and is typically used to locate and mark objects and boundaries in images. After segmentation the image represents a set of data far more suitable for further algorithmic processing and decision making, which involves tasks like locating tumors and other pathologies, measuring tissue volumes, computer-guided surgery, diagnosis and treatment planning, etc.

Over the last two decades the branch of image processing applied to medicine has evolved significantly and various publications have been presented with the goal of summarizing and evaluating this progress. Several method for assessing the quality of computer aided image segmentation (automatic or not) have been presented in (Chalana & Kim, 1997). An early work published in 1994 by Pun and Gerig (Pun, Gerig, & Ratib, 1994) has presented an outline of typical tasks involved in medical image processing, describing also common problems of such and attempts that had been taken to address them. Approaches that have been presented and discussed include the processing pipeline, image pre-processing (Lutz, Pun, & Pellegrini, 1991), filtering (Frank, Verschoor, & Boublik, 1981), early attempts of image segmentation by edge detection (Margaret, 1992; ter Haar Romeny, Florack, Koenderink, & Viergever, 1991) and region extraction (Jain & Farrokhnia, 1990; Mallat, 1989; Ohanian & Dubes, 1992), matching (Andr, Gu, ziec, & Nicholas, 1994; D. Louis Collins, Terence, Weiqian, & Alan, 1992) and recognition (Kippenhan, Barker, Pascal, Nagel, & Duara, 1992; Pun, Hochstrasser, Appel, Funk, & Villars-Augsburger, 1988). Similar work, published considerably later, has been presented in (James & Nicholas, 2000) by James and Nicholas. Authors have described accurately each step of the segmentation process, with its own difficulties and challenges and with various attempts undertaken by respective researchers to overcome them. They also elaborated on key challenges that are still to be overcame and new possible application areas for the field of computer vision. The document has been structured chronologically and researched efforts characteristic to given time period have been described. Those include in particular: era of pattern recognition and analysis of 2d images until 1984 (Alberto, 1976; Yachida, Ikeda, & Tsuji, 1980), influence of knowledge-based approaches in 1985 – 1991(Carlson & Ortendahl, 1987; Kass, Witkin, & Terzopoulos, 1988) and the development of 3D imaging and integrated analysis in later years, which incorporated more specifically: image segmentation (Chakraborty, Staib, & Duncan, 1996; Malladi, Sethian, & Vemuri, 1995; Staib & Duncan, 1996; Székely, Kelemen, Brechbühler, & Gerig, 1995), image registration, analysis of structure and morphology, analysis of function (including motion and deformation) and physics-based models. In a recent publication by Withey and Koles (Withey & Koles, 2007) the authors have presented their classification of most important medical image segmentation methods in three generations, each showing a significant level of advance comparing to its predecessor. The first generation encapsulated the earliest and lowest-level methods, including very little or none prior information. Algorithms based on image models, optimization methods, and uncertainty models composed the second generation. The third one surrounded in general the algorithms capable of incorporating knowledge.