วีดีโอ

Hi! If you want to expand (dilate) the segmentations, I do not know an easy way to do it, but you can expand annotations. you would have to convert the detections into anotations with a script: def detections = getDetectionObjects() def newAnnotations = detections.collect { return PathObjects.createAnnotationObject(it.getROI(), it.getPathClass()) } removeObjects(detections, true) addObjects(newAnnotations) and then you can expand the annotations by using the command: Objects -> Annotations... Expand annotations. If you just want to visually thicken the segmented objects, I do not know how to do it. Hope this will help!

Glad to see it's helpful!

Hi! Glad to see it's helpful, unfortunately, this image is not available, sorry.

Glad to see it's helpful!

Hi @magdalesinski3840! It's a bit hard to answer remotely but according to the error message, it seems that you did not create a project to start with. QuPath is designed to be used with projects, you can see how to create one here: th-cam.com/video/vr9w_LYtSso/w-d-xo.html. Hopefully this will resolve your problem!

Hi! Normally, when you run Cellpose the first time, it's gonna download automatically the pre-trained model you want to use. If it does not work, you need to open an anaconda prompt, activate the virtual environment and then copy the command line (from the script editor inQuPath under the script or from the console in Fiji) just above "This command should run directly if copy-pasted into your shell" in the anaconda prompt, that should run fine and download the pretrained models if downloading failed previously. I hope this will help!

Hi Ramish! It probably is the case becaus only nuclei were segmented and no extension was processed for a fake cytoplasm. In this cas, change DAB:Nucleus:Mean with DAB:Mean and that should be fine.

@@thierrypecot8747 thank you for your reply.. there are no such options in my case, I mean, it shows me very few options like area, length and related to blue , red and green .. other than these few option , in my measuring tab, no such options present as shown in video..

It means that the image is not defined as H-DAB. When you drag the image in QuPath to open it, make sure you select "Bright field (H-DAB)" in the "Set image type" drop down menu. Hope this will help!

Content de voir que c'est utile ! Il n'y a pas de méthode simple pour faire ce genre d'analyse, il serait possible d'entraîner un modèle de deep learning pour le faire, mais ce serait plus lourd. Cette méthode est probablement la plus accessible pour le faire.

Glad to see it's helpful! It's a bit difficult to answer without your images but I can't try to make some assumptions: 1/ nuclei showing high intensity for DAB might be ignored if intensity is saturating all over the nuclei, that would correspond to a flat region and might be ignored as images used for training do not show saturation; 2/ if you have both high intensity for DAB and hematoxylin and you do not expect that, it is then a staining problem, you should talk with the people at the shared facility that stain your samples. Additionally, in a normal setting, stardist is used to segment nuclei so you should not obtain non nuclear detections. If these are artefacts coming from staining and you want to disregard them, you could train an object classifier to identify "correct" nuclei and "false" nuclei. I hope this will help!

Hi! Glad to see it's helpful. I actually don't use CellProfiler so I don't plan to do it in a near future. You might have seen it, there's a tutorial available here: forum.image.sc/t/new-cellprofiler-4-plugin-runcellpose/56858 I hope this will be helpful!

Hi! Once the environment is well installed, you can run QuPath forever. However, whenever you change QuPath version, you'll have to install again the Cellpose extension and define again the environment path in QuPath. The last version changed and the env installation is not available any more, I'm gonna have a look at it to see if it's still possible to use the same virtual environment.

@@thierrypecot8747 Thank you for your reply! Yes, I have noticed that last version of CellPose extension has something different in QuPath Preferences.

@@thierrypecot8747 Hi, Thierry! Are you going to make an updated tutorial with the last versions of Qupath and CellPose extension?

Hi @@AlessioTorcinaro! I won't make a new one in a near future, but I think it's the change is pretty straight forward. If you're stuck, don't hesitate to send me a message, I'll help you then.

Hi! The best way is probably to define an annotation that only includes the regions in the slide that you want to consider for nuclei segmentation. It is also possible to change the annotation by removing the unwanted regions with alt key. The cells are still segmented but if you look at the number of cells in the annotation, it should have changed as the unwanted cells are no longer part of it. Finally, it is possible to remove cells by selecting them (double right click) and hit the suppr key. Hope this will be useful!

Hi, glad to see it's helpful! If you want to modify the classifier with a different image, it is possible to do it if you use the same project. When you go to Classify -> Object classification -> Train object classification, you open the window for the object classifier. In this window, you can click on Load training, which lets you choose which image(s) in your project you want to load your object classifier with. In that case, all the cells that were annotated in the image(s) you chose will be taken into account in your classifier, and you can add new annotations to keep training your classifier.

@@thierrypecot8747 thanks for the quick reply. can this also be applied to pixel classification?

@@moyofeyide7490 Yes, same thing

Hi! This workflow is designed to use cellpose. If you want to use Stardist, it's totally doable, you can check this video that shows how to use Stardist to segment cells with QuPath: th-cam.com/video/GBFBVT2stMQ/w-d-xo.html

Glad to see it's helpful!

Hi! There is no pre-trained cellpose model for H&E staining. You could use the nuclei pre-trained model but you'll have to select one channel. The best way to do it with QuPath is to use the "Estimate stain vectors" functionality to separate hematoxylin from eosin and then apply the cellpose model on the estimated hematoxylin channel. Hope this will help!

Hi - very interesting video, thank you. I tried to build a model on HE stained slides but it wasn't perfect. Although it was good enough for my analysis. Estimate stain vector sounds like a good option - I will give it a try @@thierrypecot8747

You're very welcome!

Glad to see it's helpful!

Glad to see it's helpful!

Hi! Sorry if the sound quality is not good enough, I'm working on improving it for the next videos. Best

Hi again Alessio! Ctrl + z works for me, but I'm still using version 0.3.0 (I'm waiting for Stardist extension to work to move to 0.4.0). However, I never found a way to remove some of the selected dots, maybe in the new version :)

@@thierrypecot8747 Yeah, same problem ;) I hope that Stardist and Cellpose, as well as trained models will work soon on v0.4.0 Thank you again

Hi Alessio! Glad to see these videos are useful. I don't know how to script auto brightness/contrast, but if you check the box "Keep settings" in the "Brightness & contrast window" below the list of channels and then apply auto contrast to a channel, it should propagate to the other images in your project. Hope this will help!

Hi Carlos! You can use a video tif, that will work for sure. Hope this will help!

Hi! You can use the dataset I put on the github (github.com/tpecot/NucleiSegmentationAndMarkerIDentification/tree/master/MaskRCNN/datasets/nucleiSegmentation_E2Fs). There is a huge dataset for fluorescence (TissueNet) available at datasets.deepcell.org/ and a large dataset for H&E images at nucleisegmentationbenchmark.weebly.com/dataset.html. There are others you can look for on the Internet. Hope this will help!

@@thierrypecot8747 thank you

Glad to see this is helpful!

Hi! It's hard to tell like this, maybe you can try on another computer. If it works then, it means that there is a problem with the installatin of Fiji on your computer, I'm not sure I would have a solution then. Good luck!

This is extremely late but in case it helps anywhere else this is an issue with newer versions of fiji! Either use an older version, or when you get the results click the "log" checkbox in the bottom right corner.

Hi John! The training and running Jupyter notebooks shown in this video for Cellpose are available at github.com/tpecot/NucleiSegmentationAndMarkerIDentification/tree/master/Cellpose. Hope this will help!

Hi Goran, glad to see this video is helpful! It is totally possible to correct annotations. If the results are not too far from what you would expect, you can just modify the annotated regions, otherwise you can remove the estimated annotations and define new ones. if you have several slides for which the staining makes the classifier fail, you can also train another classifier just for this batch. In any case, it's extremely important that you describe in the "Materials and Methods" section of the publication what you do, manual correction or training of several classifiers, and why you do it, for instance staining was different between batches of slides. Hope this will help!

@@thierrypecot8747 Great, thanks! Actually, training a different classifier is easier and more reliable than manually correcting everything. The tutorial really helped me out!

@@gorantomic6049 That's great, good luck for your research!

Bonjour Clément, Il est possible que le mode live ne soit pas suffisant. Le mieux quand on quantifie des surfaces est d'entraîner un classifieur, de le sauver puis de l'appliquer aux régions d'intérêt. Dans ce cas, les mesures doivent s'effectuer normalement. J'espère que ce sera utile.

@@thierrypecot8747 Bonjour, merci pour cette réponse. J'utilise ce logiciel depuis seulement 1 semaine et n'ayant que vos vidéos pour comprendre son fonctionnement je ne comprend pas encore très bien comment l'utiliser. Quand vous dites "entraîner un classifieur", il ne s'agit pas de sélectionner différentes zones pour avoir des annotations négatives et positives ? On peut obtenir les pourcentages sans utiliser le live prediction ? Est-il possible de discuter avec vous par mail, téléphone ou Teams ? Car j'aurais plusieurs questions a vous poser sur le logiciel. Merci d'avance

@@clementtetaud4374 Bonjour Clément, Je suis en vacances, je suis désolé, je n'ai pas pu répondre plus rapidement. Entraîner un classifieur revient à annoter des régions pour les différentes classes à estimer, dans ce cas région positive et région négative. Cela se fait dans QuPath avec "Train pixel classifier", les régions annotées seront alors considérées pour entraîner un classifieur. Le "live prediction" permet de voir l'influence de l'ajout de nouvelles annotations. Une fois le classifieur entraîné, il est possible de le sauver pour pouvoir ensuite l'appliquer à d'autres images, ce qui permet d'obtenir les mesures souhaitées, dans ce cas le pourcentage de régions positives et négatives. J'ai créé ces vidéos dans le cadre d'une formation que je fais sur le campus à Rennes, ça permet aux personnes qui assistent à la formation de revoir les différentes étapes abordées pendant la formation. Je n'ai malheureusement pas le temps de faire du conseil en dehors des collaborations que j'ai dans le cadre de mon poste à Rennes 1, je ne pourrai donc pas prendre le temps de discuter par mail, téléphone ou Teams. Le mieux est de chercher autour des vous des experts en analyse d'images. Si vous êtes dans une université, il est fort probable qu'il y en ait. Sinon, d'autres vidéos sont disponibles. La documentation de QuPath (qupath.readthedocs.io/en/stable/) est également très bien faite. J'espère que cela vous aidera dans votre recherche.

Bonjour Jack ! Je vais essayer de répondre rapidement à ces quatre points : - une fois le classifieur entraîné, il suffit de l'appliquer à de nouvelles images pour directement extraire les régions d'intérêt. Il est possible de créer un script qui va appliquer un classifieur entraîné à une collection d'images définies dans un même projet. - si le marquage est trop différent entre les images, ça peut évidemment être un problème, je conseillerais dans ce cas de discuter avec la plateforme d'histo pour voir s'il est possible d'uniformiser les techniques de marquage afin de minimiser les différences entre lames. - il est vrai que seul le CPU est utilisé, toutefois le temps de traitement par lame une fois un classifieur entraîné ne prend pas sensiblement plus de temps que l'acquisition d'une lame, donc le pipeline reste tout à fait envisageable et clairement plus rapide qu'à l'oeil, surtout que les machines peuvent aussi travailler la nuit et le weekend. - le rapport ressources / bénéfices dépend évidemment de l'application. J'espère que ces quelques remarques seront utiles.

Hi Amit. Exporting results associated with spots, you obtain a file with one line per detected object. For each line, you have the measurements associated to the given object, such as coordinates, track id ... Exporting results associated with tracks, you obtain a file with one line per track. For each track, you also have a number of measurements, such as number of spots in the track, duration, ... If you want to get the trajectory of a given object, one possibility is to export results associated with spots, then to sort the results according to the track ids (if you have several tracks) and then according to the frame, so you should end up with a file that gives you, for a given track, the successive coordinates of the tracked object. Hope it'll help.

@@thierrypecot8747 thank you very very much for the clear cut explanation 🙏🏻🙏🏻 it will definitely help a lot

Bonjour ! Évidemment, il serait possible d'ajouter les nouvelles images au même projet pour appliquer le classifieur. Il est aussi possible de récupérer le classifieur dans un autre projet. Dans ce cas, il faut aller dans le répertoire du projet dans lequel le classifieur a été créé, puis naviguer dans "classifiers/pixel_classifiers/", copier le fichier json avec le nom du classifieur (epistroma.json si le classifieur a été enregistré come epistroma), et aller le coller dans le répertoire "classifiers/pixel_classifiers/" du nouveau projet. Si ce répertoire n'existe pas, il faut le créer, et normalement il sera possible d'utiliser "load pixel classifier" et de voir le classifieur pour l'appliquer aux nouvelles images. J'espère que ce sera utile !

@@thierrypecot8747 je vous remercie!

Glad to see it's helpful!

Hi! Glad to see this video is helpful. Speed information is available in the "Edges" tab in the track table. It gives you the instantaneous speed for a given tracked object between two frames. Hope it'll be helpful.

Hi Razia! It's not surprising to get different values across the ROIs in your images, there are always variabilities when observing biological processes. However, if you have a sufficiently high number of observations, you should be able to conclude. Then, the threshold values are used to segment your objects, it is acceptable to have different values for these thresholds depending on the intensity variations that you observe locally. You can also investigate other ways to segment your objects. For instance, it might be a good idea to normalize the intensity so you can use the same threshold. At the end of the day, as for any image analysis task, there are many ways to do it, but you need to make sure you are careful, that your analysis makes sense and is well described in the method section of the articles you write. It is a bit difficult for me to offer more help, I would suggest if it's possible to find image analysis experts around you so that you can talk about it in more details. Hope it'll help.

@@thierrypecot8747 Thank you for your reply. My problem is JaCoP plugin is not handling the ROIs. I get similar values for all the ROI. In other words, ROIs are not being selected. Is there a way I can use JaCOP on images with ROIs (all ROIs were added to ROI manager). I don't know if I'm missing something in the steps to be performed. Thanks in advance

​@@raziasultana8103 Hi Razia! You're right, JaCOP does not handle ROIs, so the whole image is considered. If you want to compute the Manders coefficients in some parts of your image only, you'll have to crop your image when using JaCOP, that's the only way unfortunately. Hope it'll help.

@@thierrypecot8747 I got it.. Thank you so much for your prompt replies. I appreciate it

@@raziasultana8103 Glad to be helpful, good luck with your analysis!!!