Polyp Classification Based on Bag of Features and Saliency in Wireless Capsule Endoscopy

Wireless capsule endoscopy (WCE) enables non-invasive visual inspection of the patients' digestive tract. However, the huge number of images from the WCE has been a hurdle for doctors to handle and thus it is urgent to develop computer-aided diagnosis systems to identify problematic images. To tackle this problem, an innovative algorithm based on the integration of the Bag of Features (BoF) method and the saliency map is proposed to detect polyps from the WCE images in this study. The algorithm constitutes of four steps. In the first step, by applying the BoF method, the visual words of all images are calculated by inputting the extracted Scale Invariant Feature Transformation (SIFT) feature vectors to the K-means clustering procedure. Then we calculate the saliency and non-saliency maps of the WCE images. Following that, the histogram of the visual words of each image is calculated by integrating histograms in both saliency and non-saliency maps with various weights to represent the WCE image. Finally, polyp classification of the WCE images is conducted by Support Vector Machine (SVM) classifier. Experiments on 436 polyp images and 436 normal images are carried out to validate the proposed algorithm. The proposed method with the weight 0.9 on the saliency region achieves a best polyp detection accuracy of 92%, sensitivity of 87.9% and specificity of 93%, demonstrating that the proposed method provides a good characterization and description for polyp classification.