An ensemble of convolutional neural networks for the use in video endoscopy

In this study, a technology for creating a classifier able to identify pathological formations in images obtained with video endoscopy using the methods of deep learning is proposed. For the training and testing of neural network models, images from the CVC-ColonDB open database and 20 colonoscopy video records from the University of Arizona (Phoenix, USA) were used. To improve the performance of the proposed classification model, noise effects inherent to video cameras were considered. In addition, a study on building the model using small data samples was conducted. In building the classifier, we utilized the results of recent studies on convolutional neural networks used in medical diagnostics, which allows us to apply the proposed approach to designing the architecture of a convolutional neural network adapted to a given task. By generalizing the features of the successful models, we developed an approach towards creating a non-excessive convolutional neural network. According to the proposed approach, the network architecture is divided into blocks, which alternate to enable composing the most efficient architecture. Using the proposed approach based on the recommended selection strategy and then ranking the most significant parameters, a second approach towards building an adaptive model of classifier has been proposed. It is based on the formation of an ensemble of classifiers such as the “convolutional neural network”. To ensure the stability of the model and its insensitivity to changes in the input data as well as its applicability to different classification tasks, a set of networks with different major parameters are incorporated into the ensemble. Our experimental studies have shown that the proposed classifier can be improved by developing an ensemble of convolutional neural networks, which considers the functions proposed in the present approach. The results imply the prospective application of the developed approach for building classification models not only for medical diagnostics but also for general problems of machine vision based on small samples.