Fiji plugin to segment oocyte and zona pellucida contours from transmitted light images and extract hundreds of morphological features to describe numerically the oocyte.
Oocytor is a Fiji plugin (developped in Java). It offers the possibility to segment oocyte contours from transmitted light images, based on U-Net like neural networks which were trained on both mouse and human oocytes (in prophase and meiosis I) acquired in different conditions for robutness. It can also extract hundreds morphological features to characterize numerically the oocyte, e.g. its perimeter, circularity...
Oocytor uses neural networks for the segmentation. It was trained with images of individual oocytes placed in the center at different resolutions.
Hence, to use it you need to have a unique full oocyte in the image. When you have several oocytes in the same image, it is possible to separate them automatically with our individualize macro.
The plugin option Get cortex detects the oocyte contour of all images/stacks in a chosen folder and returns the oocyte contour as ROIs which are saved in a folder named contours automatically created in the given folder.
To use it, the plugin ask for 3 parameters:
smooth contour: possibility to smooth the final ROI contour, to avoid too wavy contours. The value of the parameter is the number of neighboring points that are taken into account to smooth. Default parameter (5) is low so as not to smooth too much.reach proportion: to build the contour, the binary image calculated from the output of the neural networks is analysed to find the limit between the segmented part (white) and the outside (black). The result can be refined by looking at the local intensity changes in the initial image and the limit will be updated to the highest point of intensity change in the neighborhood of the found limit. This parameter control the extend of this neighborhood. Low value means to stick to the pure network output while higher value implies sensitivity to the local intensity changes.nb networks: to increase the robutness of the results, it is possible to combine the outputs of several neural networks trained on this segmentation task and to take the common output. In the default version, we use 2 networks, but users can add their own networks or remove one.locate: option from version 0.5 of Oocytor, to locate the oocyte in the image (should contain only one) and zoom around it for better segmentation.
We tested different networks trained on either mouse and human data, or on mouse, human and sea urchin data (see our publication for more details).
In our models/cortex folder of this repository, we put the 3 different types of networks available: full has been trained with all data, init was trained on mouse and human data, and retrained was focused on sea urchin data.
The plugin option Get ZP detects the zona pellucida outside and inside contours of all images/stacks in a chosen folder and returns the contours as ROIs which are saved in a folder named contours automatically created in the given folder.
To perform the segmentation, we used neural networks trained on thausend images of mouse and human oocytes acquired in transmitted light. However it might be necessary to retrain it to adapt it to oocyte images, particularly if they are quite different from the one used for the training.
Our trained neural networks are available in this github in the models folder. They should be placed in a models folder in the Fiji directory (see Installation). They can be used as pre-trained networks to retrain it with only a few new images. New networks (retrain or new architecture network) can be used by oocytor plugin by replacing the oocyte*.zip folders by your own networks in the corresponding folders (models/cortex or models/zp). The names has to be kept the same.
Hundred features are implemented in Oocytor like the oocyte perimeter, circularity, Haralick's features to characterize the oocyte texture... A list is given in the repository Features list, and some will be added with the evolution of Oocytor.
This option of the plugin allows to automatically detect at which slice of a stack of images the oocyte enters in meiosis and break the nuclear enveloppe (NEBD). For this, it uses a neural network that was trained to recognize in each image if there is a visible nucleus or not. Then it calculates the moment that the nucleus disappear based on the probability of presence of the nucleus from the network output.
Remark: this neural network has been trained only with mouse oocyte datas and thus will most likely fail with other oocyte species if the nucleus doesn't have the same appearance. It can be used as a pretrain network to retrain it with different oocytes.
A compiled version is available in this release as a .jar file called oocytor_0.0.jar. To install it, simply put this file in the plugins folder of ImageJ/Fiji and restart Fiji.
You also need to put the neural networks for the segmentation in Fiji.
For this, copy the models folder in the main Fiji directory (or insert it in the models folder of Fiji if it exists already).
Oocytor uses other Fiji plugins which you may have to install if they are not already installed. See the following section.
You can also download the source code and compile it directly. As a lot of Fiji plugins, it is organized as a Maven project.
Oocytor uses the following Fiji plugins:
- CSBDeep_fiji: this plugin is used in the segmentation part. To install it, add it in Fiji update sites (go to
Help>Update.., clik onManage update sites, look forCSBDeepand select it) - FeatureJ: this plugin in the features part only. Add the
.jarfiles given in FeatureJ website to thepluginsfolder of Fiji.
Once installed, the plugin can be found in Fiji in the Plugins>CIRB>Oocytor menu.
We put in the macros folder of this repository several ImageJ/Fiji macros that can be usefull to automatize some parts of using Oocytor on several files for example.
Oocytor_DisplayROIs.ijmask to choose a file and open it with the corresponding segmentation ROIs if they have been done.Oocytor_DrawROI.ijmopen and draw the corresponding segmentation ROIs on one slice or on the full stack, for one image or all images of a folder. The color of the drawing (red for cortex and orange for zona pellucida) can be changed in the code of the macro.Oocytor_individualizeOocytes.ijmfind all the oocytes in a stack/images containing several ones and create new stack/images of cropped region containing only one full oocyte in the center.Oocytor_runAll.ijmperforms directly all the steps ofOocytoranalysis (cortex and zp segmentation, NEBD detection and features extraction) on a given file/directory.Oocytor_WriteOoCyteCenterPosition.ijmwrites the position of the ROIs cortex of the oocyte (found by oocytor) at all time/z points if it is stack of images for all the files of a given directory.
Other macros will be added while developping Oocytor and of course feel free to add yours.
- Movies of mouse oocytes maturation acquired in transmitted light are available on zenodo.
- Dataset of oocytes (mouse, human and sea urchin) with the corresponding segmentation of the cortex are available on zenodo.
An interpretable and versatile machine learning approach for oocyte phenotyping
Gaelle Letort, Adrien Eichmuller, Christelle Da Silva, Elvira Nikalayevich, Flora Crozet, Jeremy Salle, Nicolas Minc, Elsa Labrune, Jean-Philippe Wolf, Marie-Emilie Terret, Marie-Helene Verlhac
J Cell Sci jcs.260281. doi:10.1242/jcs.260281
This plugin was developped in the Terret/Verlhac team and the imaging facility Orion at the Centre de Recherche Interdisciplinaire en Biologie in the College de France.
Oocytor is freely available open-source under the GNU GPL v3.0 License (see License file).