Automatic detection of colorectal polyps larger than 5 mm during colonoscopy procedures using visual descriptors

摘要

New evidence suggests 25% - 26% of colon polyps may be missed during a routine colonoscopy[1, 2, 3, 4, 5]. These polyps or hyperplastic lesions are currently considered as pre-neoplastic lesions that must be detected. In this context, automatic strategies are appealing as second readers or diagnostic supporting tools. However, this task is challenging because of the huge variability and multiple sources of noise. This paper introduces a strategy for automatic detection of polyps larger than 5 mm. The underlying idea is that polyps in a sequence of frames are those locations with smaller frame-to-frame variance. The method starts by segmenting an input frame into a set of superpixels, i.e., clusters of neighbor pixels with minimal luminance variance. Each of these superpixels in characterized by a concatenated vector of 57 features collecting texture, shape, and color. A Support Vector Machine with a linear and Radial Basis Function (RBF) kernel was used as a supervised learning model. The evaluation was carried out using a set of 39 cases belonging to two datasets (6.594 frames: 3.123 with polyps and 3.471 without polyps) under a Leave-One-Out Cross Validation scheme and obtaining a 0.73 of accuracy. In addition, the data set was split into 70%-30% between train and test respectively and obtaining a 0.87 of accuracy.