This paper presents the adaptation of a diffusion neural network to generate a labeled synthetic dataset of electron microscopy of the brain. A model was trained can generate images and markup for them at the same time, which is an undoubted advantage of the chosen approach. Using the trained model, a set of labeled images was generated. The synthetic images are visually very similar to the original ones, the FID similarity metric between the synthetic and original datasets is 27.1. A simplified U-Net segmentation model trained on a mixed data set (original data + synthetic data) obtained a Dice score of 0.856 versus 0.858 on the original training set. Despite the good quality of synthetic data, their use in training the segmentation network does not improve the segmentation results.

Dataset extension, electron microscopy, diffusion neural networks

Yashin K.E.,  Lobachevsky State University of Nizhni Novgorod

Sokolov N.A.,  ,  orcid.org/0000-0002-9055-2225,  Lobachevsky State University of Nizhni Novgorod

Vasiliev Evgeni Pavlovich,  ,  orcid.org/0000-0002-7949-1919,  Lobachevsky State University of Nizhni Novgorod

Getmanskaya A.A.,  ,  orcid.org/0000-0003-3533-1734,  Lobachevsky State University of Nizhni Novgorod