Automatic labelling of HeLa "Kyoto" cells using Deep Learning tools

<p><strong>Name</strong>: Automatic labelling&nbsp;of HeLa &ldquo;Kyoto&rdquo; cells using Deep Learning tools</p> <p><strong>Data type</strong>: Microscopy images from the dataset &ldquo;<strong>HeLa &ldquo;Kyoto&rdquo; cells&nbsp;under the scope</strong>&rdquo;, Brightfield (BF), Digital Phase Contrast (DPC, either &ldquo;raw&rdquo; or &ldquo;square-rooted&rdquo;), Tubulin and H2B fluorescent channel, paired with their corresponding nuclei or cell/cyto label images.</p> <p><strong>Labels images</strong>: Labels images were generated using the script <em>&ldquo;prepare_trainingDataset_cellpose.ijm</em>&rdquo;.</p> <p>Briefly, for 5 defined time-points (1,10,50,100,150), channels of interest were duplicated, resaved and :</p> <p>-&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; nuclei label images were obtained using <a href="https://github.com/stardist/stardist">StarDist</a> on H2B channel</p> <p>-&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; cell label images were obtained using <a href="https://github.com/MouseLand/cellpose">Cellpose</a> on Tubulin and H2B channels</p> <p>A quick visual inspection of the resulting label images concluded that they were satisfying enough, despite certainly not being perfect.</p> <p>Notes :</p> <p>-&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; This labelling&nbsp;strategy:</p> <p>o&nbsp;&nbsp; will not produce 100% accurate labels, but they might be more reproducible than labels generated by humans and are (definitely) much faster to obtain.</p> <p>o&nbsp;&nbsp; is <strong>NOT a recommended way of generating labels images</strong>, but for educational purposes.</p> <p>-&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; The fluorescent channels are part of the dataset to ease the process of review of the labels and are NOT used for training. We generated the labels from the fluorescent channels to later predict labels from the BF or DPC channels only. As such, the fluorescent channels should not be &ldquo;reused&rdquo; with our labels during training.</p> <p><strong>File format</strong>: .tif (16-bit)</p> <p><strong>Image size</strong>: 540x540 (Pixel size: 0.299 nm)</p> <p>&nbsp;</p> <p><strong><em>NOTE</em></strong>: This dataset uses&nbsp;the &ldquo;HeLa &ldquo;Kyoto&rdquo; cells&nbsp;under the scope&rdquo; &nbsp;dataset (<a href="https://doi.org/10.5281/zenodo.6139958">https://doi.org/10.5281/zenodo.6139958</a>) to automatically generate annotations</p> <p><strong><em>NOTE</em></strong>: This dataset was used to train cellpose models in the following Zenodo entry&nbsp;<a href="https://doi.org/10.5281/zenodo.6140111">https://doi.org/10.5281/zenodo.6140111</a></p>