Transformer-based architecture for 2D semantic segmentation of mitochondria

Abstract

In this project, we have studied the state-of-the-art of semantic segmentation of biomedical images and compared the performance of a new Transformer-based architecture with the most used convolutional architecture for semantic segmentation of mitochondria in Electron Microscopy (EM) images. This is particularly interesting because both architectures are quite similar, with the main difference being the use of a Transformer as an encoder in one of them. For this comparison, we have adapted an existing Transformer-based architecture used in 3D medical images to perform 2D semantic segmentation, explored multiple variations in both the convolutional and Transformer parts, and finally performed a comparison between the two architectures under the same conditions. Furthermore, we also analyzed the impact of applying different self-supervised learning tasks as a pre-training strategy for the network.