Deep learning based generation of synthetic blood vessel surfaces
Manuela Daniela Danu, Cosmin-Ioan Niță, Anamaria Vizitiu, Constantin Suciu, Lucian Itu · 2019
In recent years, the medical imaging area showed a notably increased interest in Deep Learning (DL) based applications. Deep learning is a machine learning (ML) technique which learns features and tasks directly from data, trying to model human abstract thinking. Since deep learning can create features without a human intervention, it allows data scientists to use more complex sets of features in comparison with traditional machine learning approaches. In addition to this, the robustness to natural variations in the data is automatically learned and the deep learning architecture is flexible, so that the same neural network based approach can be applied to many different applications and data types. Our goal is to apply DL based techniques in the context of medical imaging with the purpose of developing a workflow for diagnosing cardiovascular pathologies or cerebral aneurysms. Since a major challenge of this approach is the lack of large training databases, in this paper we are focusing on performing data augmentation by generating realistic synthetic anatomical models of blood vessels. For this task we used geometries describing vessel-like structures and also real anatomies extracted from patients. We chose to experiment with two state of the art models: Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN). We address the problem of employing neural network based models on three dimensional surfaces. Such surfaces typically have an unstructured representation consisting of points and polygons and are not compatible with typical neural network architectures. We propose a technique based on surface voxelization which consists on representing the unstructured surface mesh as a three-dimensional image, therefore becoming inherently compatible with a standard convolutional neural network. We performed experiments on three datasets containing both two and three dimensional surfaces representing blood vessel-like structures. We show that state of the art, deep learning based generative models, are capable of generating voxelized three dimensional surfaces of high quality that are visually indistinguishable from the training samples.