Development of a three-dimensional convolutional neural network with attention for aneurysm detection

The paper considers a prototype of a three-dimensional convolutional neural network with an attention block detecting the probability of intracranial cerebral aneurysms in a single contrast computed tomography-angiography study. DICOM contrast computed tomography-angiography data with and without intracranial aneurysms were used to train the network. Metadata from the studies were not used. The data were divided into training and validation subsets in the proportion of 65% and 35%, respectively. Using Keras and Tensorflow libraries in the Python programming environment, a 192$\times$192$\times$128 three-dimensional convolutional neural network model with 4 convolutional layers, a kernel of dimension 3 and self-attention block was developed. The accuracy, precision and recall of classification on test samples reached 96%, 99% and 93% respectively that exceeded the performance of previously known neural networks.