Wireless capsule endoscopic frame classification scheme based on higher order statistics of multi-scale texture descriptors

The gastrointestinal (GI) tract is a long tube, prone to all kind of lesions. The traditional endoscopic methods do not reach the entire GI tract. Wireless capsule endoscopy is a diagnostic procedure that allows the visualization of the whole GI tract, acquiring video frames, at a rate of two frames per second, while travels through the GI tract, propelled by peristalsis. These frames possess rich information about the condition of the stomach and intestine mucosa, expressed by color and texture in these images. These vital characteristics of each frame can be extracted by color texture analysis. Since texture information is present as middle and high frequency content in the original image, two new images are synthesized from the discrete wavelet coefficients at the lowest and middle scale of a two level Discrete Wavelet Transform of the original frame. These new synthesized images contain essential texture information, at different scales, which can be extracted from statistical descriptors of the coocurrence matrices, which are second-order representations of the synthesized images that encode color and spatial relationships within the pixels of these new images. Since the human perception of texture is complex, a multiscale and multicolor process based in the analysis of the spatial color variations relationships, is proposed, as classification features. The multicolor texture information is modeled by the third order moments of the texture descriptors sampled at the different color channels. HSV color space is more related to the perceptive human characteristics, therefore it was used in the ambit of this paper. The multi-scale texture information is modeled by covariance of the texture descriptors within the same color channel of the two synthesized images, which contain texture information at different scales. The features are used in a classification scheme using a multilayer perceptron neural network. The proposed method has been applied in real data taken from several capsule endoscopic exams and reaches 94.6% of sensitivity and 93.7% specificity. These results support the feasibility of the proposed algorithm.