Polyp malignancy classification with CNN features based on Blue Laser and Linked Color Imaging

In-vivo classification of benign and pre-malignant polyps is a laborious task that requires histophatology confirmation. In an effort to improve the quality of clinical diagnosis, medical experts have come up with visual models with only limited success. In this paper, a classification approach is proposed to differentiate between polypmalignancy, using features extracted from the Global Average Pooling (GAP) layer of a pre-trained Convolutional Neural Network (CNNs) . Two recently developed endoscopic modalities are used to improve the pipeline prediction: Blue Laser Imaging (BLI) and Linked Color Imaging (LCI). Furthermore, a new strategy of per-class data augmentation is adopted to tackle the differences of unbalanced class distribution. The results are compared with a more general approach, showing how artificial examples can improve results on highly unbalanced problems. For the same reason, the combined features for each patient are extracted and trained using several machine learning classifiers without CNNs. Moreover to speed up computation, a recent GPU based Support Vector Machine (SVM) scheme is employed to substantially decrease the overload during training time. The presented methodology shows the feasibility of using the LCI and BLI techniques for automatic polypmalignancy classification and facilitates future advances to limit the need for time-consuming and costly histopathological assessment.