Abstract
Much of the scientific discoveries in cellular biology require segmentation of microscopy images obtained, a fundamental aspect of image analysis. Until recently, segmentation of volumetric electron microscopy (VEM) images has been a tedious and time-consuming process for researchers. Image segmentation in cellular biology is experiencing a paradigm shift as more and more of them are accomplished by deep learning (DL). Despite the development of numerous specialized tools, their generalisability and transferability suffer. Tools designed for volumetric segmentation are often specialised for a narrow range of biological specimens, limiting their usability or performance for different datasets. Therefore, it is imperative to develop an adaptable and easy-to-use tool for DL image segmentation for VEM images. In this thesis, I present the development of Volume Segmentation Tool (VST), a DL based tool that implements true volumetric image segmentation in VEM image stack data. This is a tool that is capable of segmenting a wide range of biological sample types. VST automates the handling of data preprocessing, data augmentation, and network building, as well as the configuration for model training, while adapting to the specific dataset. Furthermore, VST operates entirely on local hardware and provides a browser-based interface with additional features for visualizations of the networks and augmented datasets. It addressed common challenges in VEM images, such as large dataset size, instance segmentation, anisotropic voxels and imbalanced classes. Through examples from various subject areas on selected datasets, I demonstrate that VST achieves state of the art performance compared&evaluated to CDeep3D, DeepImageJ, ZeroCostDL4Mic and nnU-Net. VST is also an intuitive tool, its ease of use and accessibility are guaranteed with public releases, installers, low hardware requirement and a user-interface.