| Abstract [eng] |
Deep neural networks are a group of mathematical methods widely applicable in practice. The field of medicine is not an exception, especially medical diagnostics, where medical digital images are used for pathology diagnosis. However, this group of methods has a key flaw, which stops its wider adoptability in practice. That is being incapable of modelling the uncertainty of the forecast, which reduces the reliability of the method in practical settings. Gaussian processes are non-parametrical models, which are capable of modelling complex relationships and have the capability to evaluate the forecast uncertainties. In this paper we attempt to utilize these properties of Gaussian processes in modelling deep features, which are generated by convolutional neural networks, and modelling probability uncertainties, which arise from the stochastic nature of the data. The research object of this paper is digital eye fundus images, which are used for diagnostic of various pathologies. We demonstrate that the transition from the convolutional neural network to the Gaussian process does not decrease the accuracy of the diagnosis and allows us to evaluate the reliability of the diagnosis by using the uncertainty measures to identify difficult cases. These cases can then be referred to a specialist, rather than proceeding with automated diagnostics. |
