Medical image classification using a combination of features from convolutional neural networks

Medical image classification is an important and challenging problem, since images are usually complex, variable and the amount of data is relatively constrained. Selecting optimal sets of features and classifiers is a crucial problem in this area. In this paper it is proposed an image classification method, named Hybrid CNN Ensemble (HCNNE), based on the combination of image features extracted by convolutional neural networks (CNN) and local binary patterns (LBP). The features are subsequently used to build an ensemble of multiple classifiers. More specifically, the Euclidean distance between LBP feature vectors of each training class and the confidence of CNN features classified by support vector machines are employed to compose the input of a multilayer perceptron classifier. Finally, these features are also used as input to other classifiers to compose the final voting ensemble. This approach achieved an accuracy similar to those of other state-of-the-art methods in texture classification and showed an improvement of 10% over the previously reported identification of a group of odontogenic oral cyst histological images, at a low computational cost. Three major contributions are presented here: 1) the combination of low and high level features assigning weights based on the confidence of the features for texture recognition; 2) the combination of automatically learned deep features with LBP by a multilayer perceptron based on the feature confidences; 3) state-of-the-art results are obtained in the odontogenic cyst categorization problem.