Segmentation of medical images

Segmentation of medical images

Abstrakt

Segmentation constitutes a significant component of image analysis. As such, the procedure yields an image defining the position and shape of the area of interest, namely the monitored tissue; the shape descriptors can then be utilized in the analysis of tissue properties [1] or further processed. The techniques subsumed within the wide family of segmentation methods are commonly classified according to their principles. In this context, we can first refer to thresholding methods, which are based on analyzing image intensities (simple thresholding; Mumford and Shah [2]). These procedures, although frequently used, nevertheless often fail to ensure proper segmentation of inhomogeneous areas. Another subgroup of techniques comprises methods that exploit edge analysis (edge detectors [3], active contours [4]). In most tissues, the segmentation process is hampered by incomplete edges of the analyzed object; such incompleteness is caused by, for example, bonds between the given tissue and other soft or hard tissues. The discussed problem is suitably eliminated via hybrid procedures, which combine positive elements of region and edge segmentation methods. Tissues characterized by shape stability can be effectively investigated with model segmentation [5]. Current image segmentation methods involve components of artificial intelligence and learning through a set of training images [6]; despite being computationally intensive, these techniques ensure higher accuracy and robustness of the segmentation process. At present, the actual speed of segmentation can be smoothly increased via parallelization of the algorithms and their distribution into available graphic cards [7]. To determine the accuracy of the results, we can employ a wide range of parameters (such as Dice, Jaccard, sensitivity, or selectivity [8]) enabling us to compare the quality of the methods.

Segmentation constitutes a significant component of image analysis. As such, the procedure yields an image defining the position and shape of the area of interest, namely the monitored tissue; the shape descriptors can then be utilized in the analysis of tissue properties [1] or further processed. The techniques subsumed within the wide family of segmentation methods are commonly classified according to their principles. In this context, we can first refer to thresholding methods, which are based on analyzing image intensities (simple thresholding; Mumford and Shah [2]). These procedures, although frequently used, nevertheless often fail to ensure proper segmentation of inhomogeneous areas. Another subgroup of techniques comprises methods that exploit edge analysis (edge detectors [3], active contours [4]). In most tissues, the segmentation process is hampered by incomplete edges of the analyzed object; such incompleteness is caused by, for example, bonds between the given tissue and other soft or hard tissues. The discussed problem is suitably eliminated via hybrid procedures, which combine positive elements of region and edge segmentation methods. Tissues characterized by shape stability can be effectively investigated with model segmentation [5]. Current image segmentation methods involve components of artificial intelligence and learning through a set of training images [6]; despite being computationally intensive, these techniques ensure higher accuracy and robustness of the segmentation process. At present, the actual speed of segmentation can be smoothly increased via parallelization of the algorithms and their distribution into available graphic cards [7]. To determine the accuracy of the results, we can employ a wide range of parameters (such as Dice, Jaccard, sensitivity, or selectivity [8]) enabling us to compare the quality of the methods.

image segmentation

image segmentation

MIKULKA, J.

2016

09.09.2015

Brno

SMIT 2015 Proceedings

59

59

1