An intelligent system for automatic detection of gastrointestinal adenomas in video endoscopy

Abstract

Today 95% of all gastrointestinal carcinomas are believed to arise from adenomas. The early detection of adenomas could prevent their evolution to cancer. A novel system for the support of the detection of adenomas in gastrointestinal video endoscopy is presented. Unlike other systems, it accepts standard low-resolution video input thus requiring less computational resources and facilitating both portability and the potential to be used in telemedicine applications. It combines intelligent processing techniques of SVMs and color–texture analysis methodologies into a sound pattern recognition framework. Concerning the system's accuracy this was measured using ROC analysis and found to exceed 94%.

Introduction

Gastrointestinal neoplasms include polyps arising from the epithelial cells of the gastric and the colonic mucosa. These polyps are mainly classified into two types: adenomatous and hyperplastic polyps. Polyps of the first type, also referred to as adenomas, are usually cancer precursor lesions, whereas polyps of the second type are not considered to be premalignant. Definitive distinction between the two types requires polyp biopsy and histological examination of the tissue specimens. Although there are modern non-invasive procedures to detect polyps, such as virtual endoscopy, standard video endoscopy remains the most efficient minimally invasive procedure to detect even small-size polyps that allows biopsy and in many cases polyp resection. Today, the international consensus for the treatment of polyposis dictates removal of all polyps, regardless of the location, the size or other characteristics, in order to prevent a possible development of cancer [1], [2], [3].

During an endoscopic examination it is possible for some polyps to go undetected and evolve into malignant tumors in the following years. A reliable system that would be capable of supporting the detection of adenomas could increase the endoscopist's ability to accurately locate early stage adenomas, and could contribute to the reduction of the duration of the endoscopic procedure, which is in most cases uncomfortable for the patients. Such a system would minimize the expert's subjectivity introduced in the evaluation of the clinical characteristics of the examined tissue. Moreover, a consequent cost reduction of the operation would also be feasible, as more patients could be examined faster even by less experienced personnel.

A variety of methods have been proposed in the literature for computer-aided evaluation of gastrointestinal endoscopic images or video. First attempts include the application of edge detection methods for the detection of gastric ulcers [4], region-growing methods for the extraction of large intestinal lumen contours [5] and for the detection of abnormalities in the lower gastrointestinal tract [6].

By the end of the nineties, texture analysis methods combined with intelligent pattern classification techniques began to arise for the detection of lesions in endoscopic images. These methods were motivated by the fact that the textural characteristics of the tumorous lesions can be used for diagnosis not only microscopically [7] but also macroscopically [8]. Neural network-based grey-level texture analysis approaches of endoscopic images include the usage of texture spectrum [9], co-occurrence matrix [10], [11], Local Binary Patterns (LBP) [12] and wavelet- domain co-occurrence matrix features [13]. The latter approach has been applied for tumor detection in colonoscopic video-frame sequences in [14] and it was integrated in a versatile and standalone software system for the detection of colorectal lesions in endoscopic video-frames named CoLD [15].

Although texture has proved to be important for the characterization of colorectal lesions, it has been shown that color can be used as an additional clue for the detection of lesions in endoscopic images. Tjoa and Krishnan [16] combined texture spectrum and color histogram features for the analysis of colon status. Karkanis et al. [17] extended the concept of wavelet-domain co-occurrence matrix features for color images and proposed the Color Wavelet Covariance (CWC) features for computer-aided detection of adenomatous polyps of the colon in high-resolution endoscopic video-frames. The experimental results showed that these features lead to higher detection sensitivity than the original grey-level features and other color–texture descriptors [18]. In a later work, Zheng et al. [19] proposed a clinical decision support system based on a Bayesian fusion scheme that combines color, texture and lumen contour information for the detection of lumps and bleeding lesions in colonoscopic images. The fusion approach led to a marginal improvement of the system's sensitivity and specificity for lump detection as compared with the performance achieved only by extracting grey-level LBP histograms.

In this paper we present a novel intelligent system for automatic detection of colonic and gastric adenomas in endoscopic videos. It utilizes color–texture image features and incorporates non-linear Support Vector Machines (SVMs) to achieve improved detection accuracy compared to the linear classification scheme utilized in [17]. Moreover, we focus on the selection of a feature extraction method appropriate for the analysis of low rather than high-resolution video-frames. The advantages emanating from the adoption of such a method include processing time reduction, applicability in telemedicine and less demanding hardware requirements. The assessment of the system's performance is realized by means of Receiver Operating Characteristics (ROC), which provide more reliable estimates of accuracy compared to other measures, not deriving from ROC [20], which have been adopted in the previously cited works.

The rest of this paper consists of four sections. Section 2 describes the architecture of the proposed system. The methods investigated for the implementation of each module of the system are described in Section 3. In Section 4, we present the experimental results from the application of the proposed system for the detection of colonic and gastric adenomas, in colonoscopic and gastroscopic videos, respectively. Finally, the conclusions as well as future perspectives of this study are summarized in Section 5.