Formulation and statistical evaluation of an automated algorithm for locating small bowel tumours in wireless capsule endoscopy

Abstract

Wireless capsule endoscopy (WCE) is an imaging modality which is highly reliable in the diagnosis of small bowel tumors. But locating the frames carrying tumors manually from the lengthy WCE is cumbersome and time consuming. A simple algorithm for the automated detection of tumorous frames from WCE is proposed in this work. In the proposed algorithm, local binary pattern (LBP) of the contrast enhanced green channel is used as the textural descriptor of the WCE frames. The features employed to differentiate tumorous and non-tumorous frames are skewness (S) and kurtosis (K) of the LBP histogram. The threshold value of the features which offers the trade-off between sensitivity and specificity is identified through Receiver Operating Characteristic (ROC) curve analysis. At the optimum threshold, both the features exhibited a sensitivity of 100% and specificity of 90%. The skewness and kurtosis of the LBP computed from the enhanced green channel of tumorous and non-tumorous frames differ significantly (p « 0.05) with a p-value of 2.2 × 10−16. The proposed method is helpful to reduce the time spent by the doctors for reviewing WCE.