Unsupervised Ground Truth Generation for Automated Brain EM Image Segmentation

Abstract

Successful training of deep neural network models for Image Segmentation requires large datasets with proper ground truth annotations. In most bio-medical applications obtaining sufficiently large labelled datasets for training such networks, is a tedious task. Hence addressing this problem, we propose a simple light-weight neural network based model that generates ground truth masks of the neuronal structures of Electron Microscopy(EM) stacks training images. It is followed by image augmentation to create an extensive dataset of image-mask pairs for training the segmentation network. The proposed segmentation model is inspired by the state-of-the-art Unet++ architecture. We compare the segmentation predicts of the proposed model (unsupervised) with the manual ground truth masks to validate our results and efficiency of the model proposed. The proposed network model for unsupervised segmentation can be trained effectively with less number of train images even without the presence of proper ground truth masks. It predicts high quality segmentation outputs for the images under test with optimal time requirement(less than a second using a Google Colab Nvidia Tesla K80 GPU).