Informative and Uninformative Regions Detection in WCE Frames

Abstract

Wireless capsule endoscopy (WCE) is relatively a new device which investigates the entire gastrointestinal (GI). About 55000 frames are captured during an examination (two frames per second). Thus, it is bene icial to ind an automatic method to detect diseases frames or regions of a frame. The WCE videos have lots of uninformative regions (such as intestinal juice, bubbled, and dark regions); therefore, preprocessing is useful and necessary in diseases detection. In this paper, three practical methods are introduced to detect the informative and uninformative regions in a frame. In order to achieve this goal, morphological operations, fuzzy k-means, sigmoid function, statistic features, Gabor ilters, Fisher test, neural network, and discriminators in the HSV color space are used to detect uninformative regions (do not carry clinical information) in a frame. Our experimental results indicate that precision, sensitivity, accuracy, and speci icity are respectively 97.76%, 97.80%, 98.15%, and 98.40% in the irst method, 93.32%, 84.60%, 91.05%, and 95.67%, respectively in the second one, and 93.32%, 84.60%, 91.05%, and 95.67%, respectively in the third method.

Generic feature learning for wireless capsule endoscopy analysis

“…From computational point of view, most of the recent developments in WCE have been devoted to the design of algorithms to automatically detect different GI events or pathologies. Table 1 presents a review of applications of WCE such as bleeding [29,30,17,31], informative frames detection (visibility) [32,33,34,35,36], motility [22,21,37], ulcers [38,31], polyps [18,19,39,36] celiac disease [20] and Crohn's disease [24] during the last 3 years.…”

“…Year Applications Feature Bleeding Visibility Motility Ulcer Polyps Celiac Crohn's Eid et al [38] 2013 √ Texture Figuereido et al [29] 2013 √ Color Sanju et al [30] 2014 √ Color Fu et al [17] 2014 √ Color Yeh et al [31] 2014 √ √ Color Segui et al [32] 2012 √ Color and Texture Sun et al [33] 2013 √ Color and Texture Maghsoudi et al [34] 2014 √ Color, Shape and Texture Suenaga et al [35] 2014 √ Shape Segui et al [22] 2014 √ Shape Malagelada et al [21] 2015 √ Color, Shape and Texture Drozdzal et al [37] 2015 √ Color, and Shape Mamonov et al [18] 2013 √ Shape and Texture Yuan et al [19] 2013 √ Color, Shape and Texture Silvia et al [39] 2014 √ Shape and Texture Nawarathna et al [40] 2014 √ √ √ Texture Kumar et al [24] 2012 √ Color, Shape and Texture Ciaccio et al [20] 2013 √ Texture Zhao et al [36] 2015 √ √ Color, Shape and Texture [40] to detect several abnormalities such as erythema, polyps or ulcers. Eid et al [38] introduce the multi-resolution discrete curvelet transform [41] to detect ulcer images.…”

An experimental study of deep convolutional features for iris recognition

“…These features are mostly designed by computer vision experts, and are usually referred to as hand-crafted features. Different applications, such as medical image analysis [20]- [21], may use a very different set of hand-crafted features. Recently, feature-learning algorithms and multi-layer representation have drawn a lot of attention, most notably convolutional neural networks [22], where the image is fed directly as the input to the deep neural network and the algorithm itself finds the best set of features from the image.…”

Automatic Detection of Intestinal Content to Evaluate Visibility in Capsule Endoscopy

“…In all of the methods above, the part of supervised learning was limited to entire frames, while in some of them unsupervised segmentation algorithms are applied to detected frames to segment the area with intestinal content. The work done by Haji-Maghsoudi et al [8] attempts to directly classify regions instead of frames. They use a complex, multi-step approach and make use of two NN classifiers.…”