Fuzziness in Medical Image Processing: Representation and Models - Fuzziness and Medicine - page 174

constructing anatomical models used in treatment planning and estimation of tumor
border. However, the contrast in ultrasound images is usually low and the bound-
aries between the prostate and background are noise-corrupted and fuzzy. Speckle
noise and weak edges make the ultrasound images inherently difficult to segment.
Due to the challenging nature of ultrasound images, all methods proposed
in the literature are completely customized and incorporate specifically designed
and tuned pre-processing techniques to prepare the image for the segmentation
(generally thresholding).
The reader should bear in mind that an acceptable level of accuracy for these type
of images can only be achieved by a proper developed processing chain, incorporat-
ing specifically designed and tuned pre-processing and post-processing techniques.
In Fig. 12.6 we show 10 prostate ultrasound images and their manual segmenta-
tion.

Fig. 12.6 Set of Ultrasound prostate images with their ideal segmentation made by an expert
radiologist

In this study we have asked the expert to provide a point at the center of the
prostate, making object detection/extraction easier. The image has been then filtered
via median filter with 5 5 neighborhoods. Further, selective contrast enhancement

as described in [3] has been applied to increase the image quality.
The objective of this work is to study the influence of the lack of knowledge
(ignorance) of the experts with respect to the process of image segmentation. In
classical fuzzy thresholding algorithms, the expert uses a single membership func-
tion to represent the whole image (see [8, 13, 25, 26, 37]). In our proposal, the expert
should pick two different functions, one representing the background and another
one representing the object. This representation allows the expert to better model
the pixels for which he is not sure if they belong to the object or the background.
In Fig. 12.7 we show the histogram of an image and two membership functions,
one to represent the background and the other to represent the object. From these
fuzzy sets, in our proposal we introduce the concept of ignorance function G u .Such

×