Automatic detection of gastrointestinal diseases from capsule endoscopy video based on probabilistic modeling of least square saliency transformed images

Abstract

Wireless capsule endoscopy (WCE) is an effective video technology to diagnose gastrointestinal (GI) diseases, such as bleeding, ulcer, and tumor. To avoid tedious and risky manual review process of long duration WCE videos, automatic disease detection schemes are getting importance. Although many single disease classification schemes are available, a limited number of schemes to detect multiple GI diseases are found in the literature. In this thesis, at first, an automatic bleeding detection scheme is proposed based on the fitting of a characteristic probability density function (PDF) to the local features extracted from pixels of interest (POI) of a WCE image. Here, a linear separation criterion is used for POI extraction. However, POI extraction in WCE images is a difficult task due to the unavailability of sufficient number of pixel level ground-truth (PLGT) images. Even though a small number of expert-annotated PLGT images are available, they are rarely utilized for disease classification. In this thesis, we propose a least square saliency transformation (LSST) to be implemented on PLGT images to extract coefficient vectors. Next, the coefficient vectors are used to obtain transformed images with enhanced salient pixels, which offer precise POI. Later on, a characteristic PDF is fitted to the POI and the fitted PDF model parameters are used in the proposed hierarchical classification scheme. One Major advantage here is the significant reduction in feature dimension due to the use of model parameters as features, which can also capture the inherent characteristics of diseases. A disease localization method is developed by selecting the best possible transformed images with respect to the PLGT images followed by a K-nearest neighbors classification scheme. A large number of WCE images obtained from 50 publicly available videos are used for performance evaluation and the results obtained by the proposed method outperforms the results obtained by some state-of-the-art methods. As an alternate of the proposed LSST scheme, a supervised linear discriminant analysis (LDA) training scheme is also proposed that utilizes PLGT images to obtain the POI. The proposed scheme shows very satisfactory performance in classifying diseases in comparison to the LSST scheme at the expense of higher computation.