Abnormalities detection in wireless capsule endoscopy images using EM algorithm

In this paper, a novel method is proposed to detect common abnormalities in Wireless Capsule Endoscopy (WCE) video frames including Lymphoid Hyperplasia, ulcer, and angiodysplasia lesions. Inspecting WCE video frames to detect abnormality is a tedious task for physicians. One important step in the proposed approach is to extract the region of interest (ROI), i.e., suspicious region, using the expectation-maximization (EM) algorithm. Suspicious regions in WCE frames are segmented using the EM algorithm considering the color and texture information of the image. Then, suitable descriptors associated with the shape, texture, and color of ROIs are examined for further analysis. These descriptors include histogram of gradients for shape, local binary patterns for texture and different statistical characteristics from pixel values for color information. These features are then fed to a support-vector machine for classification. The results show that the proposed approach can detect abnormalities in WCE frames with the accuracy of 91.3%.