Image Processing Reference
In-Depth Information
The results received for fuzzy classification with and without the optimisation procedure
described in Section6.3.4 are listed in Table 6.5. Here we applied a genetic algorithm to
generate an optimised rule base of 100 rules; crossover probability was set to 0.9 and a
mutation probability to 0.1. It can be seen that a correct classification rate of about 80%
is achieved, which is comparable to that achieved by other techniques for breast cancer
diagnosis with mammography typically providing about 80%, ultrasonography about 70%,
MRI systems about 75% and DOBI (optical systems) reaching about 80% diagnostic accu-
racy (Zavisek and Drastich, 2005). We can therefore conclude that the presented approach
is indeed useful as an aid for diagnosis of breast cancer and should prove even more powerful
when coupled with another modality such as mammography.
goscopy
The third application we present deals with diagnosis of laryngeal pathology using im-
age data acquired from contact endoscopy. Contact endoscopy (CE) is an in vivo tech-
nique (Hamou, 1979; Wardrop, Sim, and McLaren, 2000) of obtaining detailed images of
living epithelia. It exploits a modified glass rod lens endoscope which is placed on the tissue
surface. Contact endoscopy images depict the cell organisation
(see Fig. 6.9, left column)
and the critical examination of this biological material is important to understand both
normal and pathological biological processes. In many cases, this examination depends on
the physician's level of suspicion regarding the malignancy of the lesion. In this step the
physician must visually assess various morphometric characteristics of image objects which
are visible as cell nuclei. These features help to decide if the abnormality is likely to be
malignant or benign, and to determine the recommended course of action, i.e. whether to
repeat the screening, advise a follow-up or perform a biopsy.
Until now no cell nuclei classification system in CE imaging was published. The main
problem with diagnosis using CE images is connected with the intuitive description of the
image objects attributes (see Tab.6.6). Intuitive feature description such as: “rather larger
size of the cells nucle”, “highly deformed shape of nuclei”, “high density of cell nuclei” or
“cell nuclei are grouped very closely” used in histopathological evaluations defy the precise
description of the morphometric attributes such as object size or shape coe
cient used
by the system for description of the image objects. For these reasons the intuitive expert
description can be very well modelled by fuzzy sets.
Because of the large number of input patterns and imprecise histological evaluation we
decided to apply the method of rules generation with fuzzy clustering. The aim of our
endeavours was the design and implementation of a prototype imaging system. Functional-
ity of the system is the following: (1) acquisition of image data for contact endoscopy, (2)
processing and interpretation of endoscopic image data and (3) visualisation of results. The
nuclei classification results may be treated as quantitative and qualitative additional factors
for computer aided cancer diagnosis of the larynx and may be used for selecting suitable
areas of tissue prior to biopsy. Classification results were verified by statistical analysis to
verify the potential of the proposed ideas. Computerised detection of objects (object seg-
mentation) in contact endoscopy image aim at segmenting cell nuclei or cell including the
nucleus. In this case, image segmentation can be understood as a process of grouping image
pixels into significant regions denoting the nucleus or cells. The segmentation method used
in the proposed system is described in (Tarnawski and Kurzynski, 2007) and consists of
image enhancement based on nonlinear diffusion process, followed by a modified watershed
segmentation step. The segmentation result is represented as a binary image with detected
nuclei represented by white pixels.
Search WWH ::
Custom Search