Chapaliuk B. Computer-aided diagnostic systems with artificial intelligence methods usage

Thesis research considers the problem of building computer-aided lung cancer detection systems that work with three-dimensional medical data like x-ray computed tomography. The analysis of modern artificial intelligence methods for the development of such systems is performed. Work shows that the highest diagnosis precision can be achieved using deep neural networks that take into account three-dimensional nature of the patient’s lungs image. Analysis of existing solutions and performed experiments show that the best results achieved by multi-stage three-dimensional convolutional neural networks and by recurrent neural networks. Thesis author proposed to use a combination of the two-dimensional convolution neural network and bidirectional recurrent neural network LSTM, where third spatial axis dependencies of three-dimensional data is considered by the recurrent neural network. In addition, a soft attention mechanism has been proposed to increase the learning efficiency of the proposed combined structure that allows improving the accuracy of the system by more than 8%.