UiPath manufacturing technology benefits and AI overview
T4trace master-presentation
1.
2. • Which structure can cause disorder?
• What is the most affective factor?
▲ A. J. Billnitzer, et al. (2013)
APP independent and dependent effects on neurite outgrowth
are modulated by the receptor associated protein (RAP)
3. ▲ Halavi M, Hamilton KA, Parekh R, Ascoli GA. (2012)
Digital reconstructions of neuronal morphology: three decades of research trends
4. Sholl Analysis
NeuronJ
NeuriteTracer
Method
▲ AM. Magariños et al. (2006)
Method by D. A. Sholl, (1953)
▲ E. Meijering et al. (2004)
▲ M. Pool, et al. (2008)
Approach
Counting distance boundary
Manual tracing (path finding)
Convolution skeletonize
Published
1953
2004
2008
Average time cost
-
More than 1 min.
0.3~0.5 seconds
Major Limitation
Indirect measurement
Time & handling cost
Total length/Image only
17. • Automatic thresholding: K.Srinivas, V. Srikanth, (2011) Automatic histogram threshold with fuzzy measures using C-means
• L-median skeletonize: H. Huang, et al. (2013) L1-Medial skeleton of point cloud
• Page 2. A. J. Billnitzer, et al. (2013) APP independent and dependent effects on neurite outgrowth are modulated by the receptor
associated protein (RAP)
• Page 3. M. Halavi, et al. (2012) Digital reconstructions of neuronal morphology: three decades of research trends.
• Page 4. (Left) AM. Magariños et al. (2006) Rapid and reversible changes in intrahippocampal connectivity during the course of
hibernation in European hamsters
• Page 4. (Right) M . Pool, et al., (2008) NeuriteTracer: A novel ImageJ plugin for automated quantification of neurite outgrowth
• Page 16. (Left) http://en.wikipedia.org/wiki/Sanford-Burnham_Medical_Research_Institute
• Page 16. (Right) http://www.wired.co.uk/magazine/archive/2013/09/start/clear-thinking
Good morning, I am Song yonggeunfrom Takumi Lab. I’m here to report a new method, that might beneficial to your research, that I found during my master research.
There is a study named “Neurite Morphology”. It analyze neuron shape, quantitatively, through images.As you know, neuron is primary element of mental system, and I’m sure most of you knows better, what kind of tradgedy can happen if we have deficit on it. Alzheimer, Autism spectrum disorder…Neurite morphology stands for, to figure out the mechanism, answer these questions.
The study,especially program-aided neurite morphology is now growing. Number of published papers that have used software is increasing, but I can’t think this is fast enough.There are much much more studies has neurite images, never dreamed of analyzing it.1min
The problem here, is today’s solutions are limited. Typically whenthey faced large-scale data.For example, left - traditional method “Sholl analysis” which is widely accepted until today, has risk of biases because it is indirect measurement.NeuronJ on the middle, and those relative tools, requires a lot of manual handling. You have to make trace one by one.And NeuriteTracer, quite popular today, can only show total length per an image as an output.There are many other tools, but each one of them has their own limitation like these.Most tools are fine if we have dozens of images. But what if it’s Thousands of images? What if Millions?
There I realized, A new method is urgently needed.It has to be automatic, easier to be adapted.Comprehensive, providing structural details.Reliable, the most, needless to say, the measurement have to be accurate. I started my try from this point.Classical methods had many obstacles to be automated. The term “automate”, wasn’t just meant for easy-handling.To automate the process, It had to overcome various conditions; There are many different microscopes with different settings – different resolutions, zoom levels or so.Directly to the conclusion, I got the solution and I named it “T4trace” and I’d like to show you “How”.2min.
T4trace builds reconstruct model from loaded image data, then measure the model and output.Let me explain the procedure, step-by-step.
Before begin,In a computer, an image file you usually see isn’t that visual. It is actually a sets of numbers.It is called “Bitmap data”, or “Raster Image data”.You can see it clearly, when you zoom-in closer.A cell in the grids is called a pixel. Pixel contains attributes for the area. Most important parts are color intensity. R.G.B. here.3min.
Most generally used 8-bit format, you’ll have two-to-power-of-eight (256) levels of intensity, per a color.As intensity is higher, signal is getting stronger, therefore the pixel is brighter.From a 1,000 sample images I got, I could get rough-sense on those ranges and distribution from my experience. Histogram on the left. Greens are signal, Pinks are noises here.You may noticed that It is definitely impossible to clear out all the noises completely.To solve the problem, I picked up 2011 published fuzzy threshold algorithm.It first pick a cut-off. Separate each parts, get mean expectation from both side. then take center, Move over, until it hits the value visited before.If you start from here, the value bouncing up-and-down, till it’s getting converged into proper range.On the sample set, the threshold value usually ranged here, around 80(+/-) 20, so you can see it worked.-4min.
Next step is sampling phase.As I mentioned on “Bitmap data”, the image here, is also a set of numbers.We human, can simply see, distinguish, then measure. But a machine can not make such naïve decision.So, We must convert those numbers to a computable object. What I call “Node”.One dot here indicates one node.Then bring the nodes into computation.
To build skeleton structure, I adapted 2013 published L1-medial skeletonize method, then I made some changes.Basic Idea is simple. Let’s imagine you have a loaf of bread. Then you inject special Anko, red-bean in it, that can freely move around inside of the bread.What if, those special anko, interacting, gathering into the core part of the bread? You’ll get the seam inside, like skeleton of the bread. Exactly like that.The criteria is two things: Number of neighbors in given distance boundary, and linearity.For each nodes, If there is not enough active neighbors near by, fix it. That’s done for it.You can see those color had been changed to green.If there’s enough neighbors to test, and if they were aligned in a line, the node will getting nudged into the line.else, merge neighbors to the node.Doubling the boundary, Do the same round again, until every nodes are getting fixed.5 min.
After all nodes getting fixed, we link nodes, to build a reconstruct model.One trick here is – since soma usually shows most strongest signal above all,Pick the biggest node as a soma, then expand the structure to nearbys.If you overlay the model into source image, you can see it nicely matched.6min
At this moment, let me show you a brief demo.The sample images were provided from other experiment, Primary culture, Ctxneuron from ICR E15 mice.If you double-click an image on the list, you’ll get the result instantly, in a wink.Plus, You can check procedure, with interactive image layers.And also quantified structure data.For “high-throughput”, Let me show you, a 100 image processing on the fly.As you can see here, you don’t have much parameters. Actually these are optional. You don’t really need to touch these- There you go.7min.
Now, back to the presentation, It seems nice, but are those numbers are correct?I made 100 Manual trace, take that as a control. Then compared to program tracing outputs.Say if you got 1,000 pxlength from manual trace. And you got 900 with the same source by a program, then the difference rate will be 10%.In that sense, I compared T4trace with NT.T4T were definitely more accurate NT. Plus, the variance was relatively small, which indicates it rarely goes wrong.
You can confirm the fact with correlation, too.T4trace showed 10% higher correlation than NT, Therefore, We can guarantee that T4trace is accurate – at least, better than NT.8min.
So here we found a new solution, that is automatic, comprehensive, and reliable.Addition to it, “T4trace” is also universal that can be run in most desktop-environment, I mean You can try it on your labtop, with your own data.And it is definitely fast and easy to use. I hope you to try once.
I’m expecting the solution might helphigh-throughput screening, or large-scale reconstruct.Or, Maybe It can find its usability on outside of neurite morphology, i.e. cell detection or so.
This presentation has refered many as listed here.I’d like you to focus Top two, about algorithms.Others are image refereces used on these slides.
You can freelydownload the software from the website. I hope you to try T4trace with your own data.Just before closing, I want to thank to K. Fukumoto and Takumi sensei for advisory.This concludes my talk.Thank you very much