Automated classification of nasal polyps in endoscopy video-frames using handcrafted and CNN features

Abstract

Nasal polyps are edematous polypoid masses covered by smooth, gray, shiny, soft and gelatinous mucosa. They often pose a threat for the patients to result in allergic rhinitis, sinus infections and asthma. The aim of this paper is to design a reliable rhinology assistance system for recognizing the nasal polyps in endoscopic videos. We introduce NP-80, a novel dataset that contains high-quality endoscopy video-frames of 80 participants with and without nasal polyps (NP). We benchmark vanilla machine learning and deep learning-based classifiers on the proposed dataset with respect to robustness and accuracy. We conduct a series of classification experiments and an exhaustive empirical comparison on handcrafted features (texture features-Local Binary Patterns (LBP) and shape features-Histogram of Oriented Gradients (HOG) and Convolutional Neural Network (CNN) features for recognizing nasal polyps automatically. The classification experiments are carried out by K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Random Forest (RF), Decision Tree (DT) and CNN classifiers. The best obtained precision, recall, and accuracy rates are 99%, 98%, and 98.3%, respectively. The classifier methods built with handcrafted features have shown poor recognition performance (best accuracy of %96.3) from the proposed CNN classifier (best accuracy of %98.3). The empirical results of the proposed learning techniques on NP-80 dataset are promising to support clinical decision systems. We make our dataset publicly available to encourage further research on rhinology experiments. The major research objective accomplished in this study is the creation of a high-accuracy deep learning based nasal polyps classification model using easily obtainable portable rhino fiberoscope images to be integrated into an otolaryngologist decision support system. We conclude from the research that using appropriate image processing techniques along with suitable deep learning networks allow researchers to obtain high accuracy recommendations in identifying nasal polyps. Furthermore, the results from the study encourages us to develop deep learning models for various other medical conditions.