Feature extraction and pattern classification of colorectal polyps in colonoscopic imaging

A computer-aided diagnostic system for colonoscopic imaging has been developed to classify colorectal polyps by type. The modules of the proposed system include image enhancement, feature extraction, feature selection and polyp classification. Three hundred sixty-five images (214 with hyperplastic polyps and 151 with adenomatous polyps) were collected from a branch of a medical center in central Taiwan. The raw images were enhanced by the principal component transform (PCT). The features of texture analysis, spatial domain and spectral domain were extracted from the first component of the PCT. Sequential forward selection (SFS) and sequential floating forward selection (SFFS) were used to select the input feature vectors for classification. Support vector machines (SVMs) were employed to classify the colorectal polyps by type. The classification performance was measured by the Az values of the Receiver Operating Characteristic curve. For all 180 features used as input vectors, the test data set yielded Az values of 88.7%. The Az value was increased by 2.6% (from 88.7% to 91.3%) and 4.4% (from 88.7% to 93.1%) for the features selected by the SFS and the SFFS, respectively. The SFS and the SFFS reduced the dimension of the input vector by 57.2% and 73.8%, respectively. The SFFS outperformed the SFS in both the reduction of the dimension of the feature vector and the classification performance. When the colonoscopic images were visually inspected by experienced physicians, the accuracy of detecting polyps by types was around 85%. The accuracy of the SFFS with the SVM classifier reached 96%. The classification performance of the proposed system outperformed the conventional visual inspection approach. Therefore, the proposed computer-aided system could be used to improve the quality of colorectal polyp diagnosis.