ZeroCostDL4Mic / DeepBacs - Multi-label U-Net training dataset (Bacillus subtilis) and pretrained model

<p>Training and test images of live <em>B. subtilis </em>cells expressing FtsZ-GFP for the task of segmentation.</p> <p>Additional information can be found on this <a href="https://github.com/HenriquesLab/DeepBacs/wiki">github wiki</a>.</p> <p>The example shows the fluorescence widefield image of live <em>B. subtilis </em>cells expressing FtsZ-GFP, the manually annotated instance segmentation mask and the corresponding 2-label semantic segmentation mask used for model training.</p> <p>&nbsp;</p> <p><strong>Training and test dataset</strong></p> <p><strong>Data type</strong>: Paired fluorescence and segmented mask images</p> <p><strong>Microscopy data type</strong>: 2D widefield images (fluorescence)&nbsp;</p> <p><strong>Microscope</strong>: Custom-built 100x inverted microscope bearing a 100x TIRF objective (Nikon CFI Apochromat TIRF 100XC Oil); images were captured on a Prime BSI sCMOS camera (Teledyne Photometrics)</p> <p><strong>Cell type</strong>: <em>B. subtilis</em> strain SH130 grown under agarose pads</p> <p><strong>File format</strong>: .tiff (8-bit)</p> <p><strong>Image size</strong>: 1024 x 1024 px&sup2; (Pixel size: 65 nm)</p> <p><strong>Image preprocessing</strong>: Images were denoised using PureDenoise and resulting 32-bit images were converted into 8-bit images after normalizing to 1% and 99.98% percentiles. Images were manually annotated using the Labkit Fiji plugin and mask images with labeled cytosol and cell boundaries were created using a custom Fiji macro (see our <a href="https://github.com/HenriquesLab/DeepBacs/tree/main/ImageJ-macros">github repository</a>).</p> <p>&nbsp;</p> <p><strong>Multi-label U-Net model</strong>:</p> <p>The U-Net (2D) multilabel model was generated using the ZeroCostDL4Mic platform (Chamier &amp; Laine et al., 2021). It was trained from scratch for 200 epochs on 733 paired image patches (image dimensions: (1024 x 1024 px&sup2;), patch size: (256 x 256 px&sup2;)) with a batch size of 8 and a categorical_crossentrop loss function, using the U-Net (2D) multilabel ZeroCostDL4Mic notebook (v 1) (Chamier &amp; Laine et al., 2021). Key python packages used include tensorflow (v 0.1.12), Keras (v 2.3.1), numpy (v 1.19.5), cuda (v 11.1.105). The training was accelerated using a Tesla P100GPU.</p> <p>&nbsp;</p> <p><strong>Author(s)</strong>: Mia Conduit^1,2, S&eacute;amus Holden^1,3</p> <p><strong>Contact email</strong>: <a href="mailto:Seamus.Holden@newcastle.ac.uk">Seamus.Holden@newcastle.ac.uk</a></p> <p>&nbsp;</p> <p><strong>Affiliation</strong>:</p> <p>1) Centre for Bacterial Cell Biology, Biosciences Institute, Newcastle University, NE2 4AX UK</p> <p>2) ORCID: 0000-0002-7169-907X</p> <p>&nbsp;</p> <p>&nbsp;<strong>Associated publications</strong>: Whitley <em>et al</em>., 2021, Nature Communications, https://doi.org/10.15252/embj.201696235</p>