Mean shift-based lesion detection of gastroscopic images

Abstract

Gastroscopy is one of the most important ways for diagnosing gastric cancer. Computer-aided detection of gastroscopic images is helpful in improving the accuracy of gastric cancer diagnosis. This paper proposes a method for lesion detection of gastroscopic images. Mean-shift segmentation is initially applied to reduce the information interference caused when global image or rectangular block serves as an identification area. A well performed three-dimensional color histogram feature is extracted from YCbCr color space. Mean shift-based Color Wavelet Covariance (MS-CWC) is proposed to reduce the cost of computing. Finally, after comparing Perceptron with AdaBoost, the latter is selected to train the classifier for detecting abnormal regions in gastroscopic images. Experiments show that the proposed method is feasible for lesion detection of gastroscopic images; the false negative rate(FNR), false positive rate(FPR), and error rates(ER) are 15.50%, 16.89%, and 16.35%, respectively. © 2012 Springer-Verlag.