Uninformative frame detection in colonoscopy through motion, edge and color features

Abstract

Colonoscopy is performed by using a long endoscope inserted in the colon of patients to inspect the internal mucosa. During the intervention, clinicians observe the colon under bright light to diagnose pathology and guide intervention. We are developing a computer aided system to facilitate navigation and diagnosis. One essential step is to estimate the camera pose relative to the colon from video frames. However, within every colonoscopy video is a large number of frames that provide no structural information (e.g. blurry or out of focus frames or those close to the colon wall). This hampers our camera pose estimation algorithm. To distinguish uninformative frames from informative ones, we investigated several features computed from each frame: corner and edge features matched with the previous frame, the percentage of edge pixels, and the mean and standard deviation of intensity in hue-saturation-value color space. A Random Forest classifier was used for classification. The method was validated on four colonoscopy videos that were manually classified. The resulting classification had a sensitivity of 75 % and specificity of 97 % for detecting uninformative frames. The proposed features not only compared favorably to existing techniques for detecting uninformative frames, but they also can be utilized for the camera navigation purpose.