Applying a Synthetic Dataset in Medical Images Segmentation Problem

Abstract

The paper is devoted to the problem of vessel segmentation in CT scans of human lungs. Segmentation is performed based on a machine learning model — a U-Net type convolutional neural network. To train the model, we need a set of segmented images — a training dataset. Manual segmentation of blood vessels on CT images is a laborious task. Also there are no open lungs vessels labeled or segmented datasets available at the moment. In the paper we propose an alternative approach — algorithmic generation of a vascular trees and further using of these trees as a synthetic dataset. We investigate the possibilities of training the model on such dataset and its further use for segmentation of real CT scans of the lungs.