Lineage Tracking and Asymmetric Inheritance Patterns of the Diatom Seminavis robusta [U-Net Training source code]

This repository belongs to the observation experiments for tracking lineages of Seminavis robusta. The connected work reconstructed the lineages and obtained individuals' timing of division and motility. Cells were entrained for 5 days prior under white light with 25 µmol phot / s in a 12:12 hour rhythm. The images were recorded with a modified Leica microscope (DMIL) and 770 nm light source and a band pass filter. Entrainment continued for the time of recording.
This repository contains code for training a U-Net model to segment microscopy images into three classes: cell aggregates, single cells, and background. The implementation is adapted from Lee et al. (DOI: 10.1038/s41598-022-12532-7), originally published under the MIT License. During training, the code computes standard and custom segmentation metrics at user-defined intervals (in a separate process to avoid I/O bottlenecks). After the final epoch, these metrics are automatically plotted and saved to the output directory. The provided version includes image resizing functionality and uses segmentation masks generated by tools developed for this work. Furthermore, code for testing on unseen data, all (pickled) metrics and training results, and image data are provided.
Use: Configure paths (training data/output directory), learning rate, epochs, and metric calculation frequency in config.py. Set the random seed and CUDA devices in main.py. Run main.py from IDE.
License: This repository is licensed under the European Union Public Licence v. 1.2 (EUPL-1.2). Portions of the code originating from Lee et al. are licensed under the MIT License.
In accordance with the respective licensing requirements, both license texts are included in this repository. See the LICENCE (EUPL-1.2) and LICENCE_MIT files for details.