cellfinder v1.3.0 blog post (#197)

* V1 cellfinder 1_3 blogpost

* Fixed typo

* Set plots to have 60% width

* Apply suggestions from code review

Co-authored-by: Adam Tyson <[email protected]>

* Switched from relative to absolute width for images

---------

Co-authored-by: Adam Tyson <[email protected]>

docs/source/blog/cellfinder-v1_3.md

@@ -0,0 +1,51 @@
+---
+blogpost: true
+date: June 3, 2024
+author: Igor Tatarnikov
+location: London, England
+category: brainglobe
+language: English
+---
+
+# Cellfinder version 1.3.0 is released!
+
+We are excited to announce that a new version of `cellfinder` has been released. 
+
+## Main updates
+ * This update brings a significant change to the backend of `cellfinder`, as we have switched from TensorFlow to PyTorch. This change allows `cellfinder` to support python versions 3.11+, and simplifies the installation process. The new `cellfinder` version maintains the same classification accuracy. Models trained using previous versions of `cellfinder` will continue to work with the new version.
+
+
+ * The default batch size used for detection has been increased to 64, which improves classification speed by approximately 40% on most systems. The batch size used for detection can now also be adjusted in the `napari` plugin.
+
+## What do I need to do?
+
+We recommend using a fresh conda environment to simplify the update.
+For GPU support, please follow the installation instructions in the [documentation](../documentation/setting-up/gpu.md).
+
+```bash
+conda create -n cellfinder -c conda-forge python=3.11
+conda activate cellfinder
+pip install cellfinder
+```
+
+You can also update an existing installation of `cellfinder` using pip:
+
+```bash
+pip install --upgrade cellfinder
+```
+
+
+## Classification performance
+The classification performance between the two versions is comparable. Below is a comparison of the performance between the two versions using data from the [`cellfinder` paper](https://doi.org/10.1371/journal.pcbi.1009074). Running `cellfinder` with a PyTorch backend results in a  comparable Pearson correlation and slightly improved linear best-fit slope (labelled as "coeff" in the plot) when comparing to manual cell counts. For more details on how the plots were generated, see the [`cellfinder` paper](https://doi.org/10.1371/journal.pcbi.1009074).
+
+### TensorFlow backend
+
+<p align="center">
+    <img src="../_static/comparison_tensorflow.png" width="800px">
+</p>
+
+### PyTorch backend
+
+<p align="center">
+    <img src="../_static/comparison_torch.png" width="800px-">
+</p>