Combination of Color-Based Segmentation, Markov Random Fields and Multilayer Perceptron

Abstract

Angioectasias are lesions characterized by specific features, related to the color and shape. These lesions have a cherry red appearance, due to the nature of its origin (inflammation of blood vessels); and usually are characterized with a circular shape. As was previously reported, CIELab color has high efficiency in differentiating colors in an image (Connolly and Fleiss 1997). This color space is composed of three different channels: L represents the lightness information that goes from 0 (black) to 100 (diffuse white), and the components a and b represent the color-opponent dimensions. Negative values of channel a indicate green and positive indicate magenta (adequate for the detection of red color); and negative values of b indicate blue and positive indicate yellow (Weatherall and Coombs 1992). Since the red color in these lesions is easily spotted in the middle of the gastrointestinal tissue, CIELab was the chosen color space for processing the lesions. A pre-processing was included to overcome some problems with high values of the component a that do not represent a red area on the tissue. This step was designed by observation of a high number of images, leading to the understanding that certain areas of the images (usually near bubbles) could lead the segmentation step to fail. As was already shown in a previous work of the authors (Vieira et al. 2016), these types of lesions can be separated from the normal tissue by using probabilistic segmentation methods, as the Maximum A Posteriori (MAP) approach. This method was complemented with the use of Markov Random Fields (MRF) to improve the border definition. Because the segmentation