Automatic detection of informative frames from wireless capsule endoscopy images

Wireless capsule endoscopy (WCE) is a new clinical technology permitting visualization of the small bowel, the most difficult segment of the digestive tract. The major drawback of this technology is the excessive amount of time required for video diagnosis. We therefore propose a method for generating smaller videos by detecting informative frames from original WCE videos. This method isolates useless frames that are highly contaminated by turbid fluids, faecal materials and/or residual foods. These materials and fluids are presented in a wide range of colors, from brown to yellow, and/or have bubble-like texture patterns. The detection scheme therefore consists of two steps: isolating (Step-1) highly contaminated non-bubbled (HCN) frames and (Step-2) significantly bubbled (SB) frames. Two color representations, viz., local color moments in Ohta space and the HSV color histogram, are attempted to characterize HCN frames, which are isolated by a support vector machine (SVM) classifier in Step-1. The rest of the frames go to Step-2, where a Gauss Laguerre transform (GLT) based multiresolution texture feature is used to characterize the bubble structures in WCE frames. GLT uses Laguerre Gauss circular harmonic functions (LG-CHFs) to decompose WCE images into multiresolution components. An automatic method of segmentation was designed to extract bubbled regions from grayscale versions of the color images based on the local absolute energies of their CHF responses. The final informative frames were detected by using a threshold on the segmented regions. An automatic procedure for selecting features based on analyzing the consistency of the energy-contrast map is also proposed. Three experiments, two of which use 14,841 and 37,100 frames from three videos and the rest uses 66,582 frames from six videos, were conducted for justifying the proposed method. The two combinations of the proposed color and texture features showed excellent average detection accuracies (86.42% and 84.45%) with the final experiment, when compared with the same color features followed by conventional Gabor-based (78.18% and 76.29%) and discrete wavelet-based (65.43% and 63.83%) texture features. Although intra-video training-testing cases are typical choices for supervised classification in Step-1, combining a suitable number of training sets using a subset of the input videos was shown to be possible. This mixing not only reduced computation costs but also produced better detection accuracies by minimizing visual-selection errors, especially when processing large numbers of WCE videos.