Convolutional neural network for blood vessel segmentation : 2D and 3D architectures

Details

Supervisors

De Vleeschouwer, Christophe ; Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

Recently, thanks to the success of deep learning models in image processing and computer vision, many image segmentation techniques using deep learning have emerged. Some of these models have a huge potential for medical imaging. In particular, convolutional neural networks are very efficient for the segmentation of blood vessels allowing to detect vascular diseases. In this work we propose convolutional neural networks based on the well-known architecture U-Net which are intended to segment blood vessels in 3D volumes. Among the great challenges in the processing of 3D datasets by full 3D convolutional networks, we find the slow execution time and the high memory requirements. To try to remedy these problems, a method modifying the convolution operations performed in convolutional neural networks by using 2D orthogonal cross-hair filters is presented. This method, which depending on datasets and network configurations, can save time, memory or improve the accuracy of the predictions or at least obtained results very similar to those obtained with classic convolutions, while using 3D context information.