Machine Learning in ImageJ/Fiji - StarDist transcript

hi welcome to the channel earlier on in the year i made a video on the plug-in the machine learning plug-in for slicer 3d but i think that most people who watch this channel are more interested in image j and fiji now i think all this machine learning and artificial intelligence is an absolute game changer completely different way to do image analysis of course it's quite complicated it requires you perhaps to to train an algorithm and then apply it in some way but fortunately there are some options in mhj and fiji that allows to do this in a relatively more simple way although there is still a way to train your own algorithm i'm going to try and learn how to to do that and maybe you like might like to join me on this little journey or a road to machine learning if you like the first one i'm going to look at is stardust this is a trained algorithm that will help you to segment nuclei fluorescent nuclei now i study blood vessels and i look at the nuclei and blood vessels all the time so i think this stardust could be extremely useful for me in later videos what i intend to do is try to train my own algorithm and i'll show you how i do it i don't know anything about um python or c sharp or really anything about computing science in that sense but i think it's possible to to at least make some sort of progress in this field um with what's available in msg and fiji so well let's just kick off see how we get on okay so i've got a z series it's a blood vessel um it's got lovely nuclei here the endothelial cell nuclei on the inside of the blood vessel here are the smooth muscle cell nuclei fluorescent stain and here's the adventitial smooth muscle that's sorry some added tissue cells on the outside so adventitia smooth muscle endothelium the nice thing about these nuclei is that they have quite distinct shapes um as you can quite clearly see all right so normally when we would be thresholding these we would call image adjust threshold and you can see you can do quite a good job of thresholding although there's a bit of a clump of nuclei there which clearly are individual nuclei but uh yeah but it's seen as one clump uh okay but we don't want to do the adjust the analyze particles and so on want to see if the stardust machine learning plugin will give us a better result first thing you need to do is to go to the plugin page so do a search for stardust mhj plugin and you will arrive here imagej.net stardust you get some information about the plugin the people who wrote the plugin um you get some information here about the parameters that you can adjust and what they do there's also some information up here about how you would go around training your own algorithm which is something that i i hope to do in the very near future i'll make a video on that too so okay find the plugin install it in the normal fashion for uh fiji and mhj the instructions are all there right so we'll go to the plugin let's let's start with let's make sure you get that image selected so we're doing the endothelial one um right let's go plugins right down the bottom this has come off the screen grab area but i'm i'm looking at the stardust plugin down here choose a stardust and it pops up this window here there are quite a few trained algorithms that you can choose i'm just going to go for fluorescent versatile fluorescent nuclei because well that's what i've got is a fluorescent nuclei i won't change any of these parameters let's just um yeah let's just click ok see what happens was quick wasn't it wow wow that is impressive isn't it i tell you what's most impressive see this little clump of nuclei here it's these guys there right now you know if you were to adjust threshold yeah the old-fashioned way it's going to be really difficult to to segment out those those three nuclei but but check that out wow that's impressive um yeah so that was that was just the default let's try the let's try the adventitial cells i'm going to go back well one thing i should show you is that what it does is that they're all different colors because they're all different individual objects so select them all as individual objects fabulous let's do the adventitial cells see how they do plug in stardust 2d here we go not changing any parameters not too bad at all it's maybe picked up quite a few here that are not actually nuclei so maybe we could play with the parameters a little bit uh should really check the instructions i can't remember now which what slider it was to reduce the numbers i decided it was either percentage lower percentage high okay so previously we had it 0.5 and 95 let's uh go to nearly 25 see what happens it looks as if it's actually made more nuclei so let's see if we can make it do less which would presumably be then taking that back to zero maybe taking this down a little bit yes a little bit better isn't it still got some nuclei here which don't actually exist but yeah i think that's uh it's pretty impressive for for a set of images that it hasn't been trained on and remember it's important to remember that the algorithm was not trained on these type of nuclei now it certainly wasn't strained on these type of nuclei these long thin nuclei so let's see how it does with that go stardust just go back to the defaults [Music] and see how we do well not bad but as you can see what it's done is it's detected the nuclei quite nicely but it split them up into up individual objects so you can see here that here's this nice long nucleus here which is probably can't find it maybe this one here i think i know but anyway and you can see one two three four it looks like it's it thinks it's four different objects but that's presumably because it's been trained to find smaller units but i think you'll agree the segmentation is really good it's just it's just how it classifies the segmentation okay that's probably all for now all i wanted to show in this short video was that just by installing the stardust plugin and running it on fluorescent nuclei seems to work extremely well for some nuclei works particularly well for the endothelial cell nuclei and those nuclei i suppose must be similar in size to the training data set i think what i might need to do is to train an algorithm to identify smooth muscle cell nuclei that could be interesting watch this space okay i hope you enjoyed that um or at least maybe give you the confidence to have a goal uh um trying some of this machine learning um on your own images good luck if you liked what you saw then please give a thumbs up subscribe all that stuff and i'll see you next time