Abstract:
This work is a result of discernment of the fact that the implementation of computer
aided design (CAD) for the diagnosis of cancer in gastrointestinal (GI) tract
has prospective towards a vigorous effort for the betterment of mankind. There exists
multiple levels of relevant literature and research for the configuration of CAD
system. We aimed to unfold the contribution of a computer vision module in the prototype
of this system. The basic infrastructure of the this module finds its strongest
roots to rely on the layout of pattern recognition; which encompasses classification
of images all the way from image restoration, to segmentation followed by feature
extraction. We have studied the general characteristics of endoscopic images to
comprehend the requisites, constraints and challenges towards the development of
this module. Later we made our research more specific by particularly focusing on
Chromoendoscopy imaging modality for GI tract. Clinically pertinent images were
filtered out from the pool of images, recorded during live endoscopic scrutiny. We
have devised a novel method for color feature extraction using adapted color space
and exploiting classical bag-of-words approach. Aware of the fact, that the dynamic
imaging conditions are inevitable, we encountered the issue of illumination variation
in the images; that was tackled using homomorphic filtering. Later, we integrated
color features with texture using Homogenous Texture descriptor based on the Gabor
filter responses. We, then verified the research with results affirming the claims
to be generic, robust and reliable. In order to prove the better performance of the
proposed descriptor, a comparison between different descriptors was drawn.
