building empathic. fellow @thielfellowship, prev cs@stanford
San Francisco, CA
- Tweets 195
- Following 942
- Followers 3,176
- Likes 503
Labeling bio images for AI just got 10x faster🏎. Today, we’re releasing Biodock Autolabel, and it’s kind of magical. We trained it on thousands of bio images so you can use it without training. (1/6)
@BiodockAI #microscopy #pathology #biology
36
411
2,164
Labeling bio images just got 10x faster 🏎️
Watch our best AI model yet in action, released today as AI Detect on @BiodockAI.
You prompt one object, it finds the rest. Try it for free now!
31
267
1,400
205,845
Today, we’re announcing that Biodock is coming out of beta. Biodock is the easiest way for biologists to create a remarkably accurate, AI pipeline for any bio images 🔬. It takes an afternoon, doesn’t require any code, and is incredibly accurate.
A thread (1/8):
8
180
1,086
Introducing AI Click powered by Segment Anything (SAM) for biological images, on Biodock. We’ve worked hard to make SAM much more useful for bio - read below to find out how!
9
87
406
57,577
Introducing AI Prelabel for all bioimages 🔬 - build AI models using AI models. Read about how it works below! (1/10)
@BiodockAI #microscopy #pathology #Biology
2
45
165
32,017
Try it for free on Biodock at app.biodock.ai by starting a new AI project and selecting Autolabel. For R&D customers, we can even host your own private Autolabel trained on your use case. Docs at docs.biodock.ai (6/6)
3
7
57
In just a few hours, you could build a custom AI pipeline trained on your images to:
- Quantify the morphology of different organelles
- Accurately segment neuronal dendrites and soma
- Segment out regions of an H&E-stained lymph node
and much more!
(2/8)
2
1
45
Calling bioinformaticians👨💻: Announcing the Biodock API - integrate AI image analysis at scale with a POST request, now in public beta and free to use! (1/6 🧵)
#bioinformatics #microscopy #pathology #biology
2
9
41
5,108
Biological images often contain objects of dramatically different scales.🔍
Introducing Size Groups on @BiodockAI, which allows AI to zoom in and out like a telescoping lens to see your biological objects at the optimal scale. 🧵 (1/5)
#microscopy #pathology #biology
1
3
35
4,959
If you don’t have time to label, we have labelers who can help you. If you’re an academic, everything is totally free to use. And if you are from a commercial organization, you’ll see exactly how accurate your AI pipeline is on your images before you pay a dime. (7/8)
1
1
31
We’ve seen stellar teams across academia and industry bogged down by image analysis. They often circle by hand, use off-the-shelf software that barely works on their images, or even abandon their image assay because they can’t get something that works consistently.
(3/8)
1
2
28
We’re looking to change that - we think that if you can see something in your images of cells/organoids/tissues/bacteria/other, you should be able to quantify it across groups with a high degree of accuracy. So how does it work?
(4/8)
1
27
Labeling objects is one of the most painful parts of training AI, especially in bio. Clicking 40 times or dragging the pen tool to define an ROI for hundreds of objects takes a very long time. And yet precise labels are so important for AI accuracy. (2/6)
1
1
27
Just upload your files into Biodock, label with our image labeler built for biological images, and click train. We train a custom state of the art AI model for you optimized for bio images and drop it into your account as a pipeline you can run in one click.
(5/8)
1
1
25
Your new pipeline:
- Outputs directly to powerful graphs, statistics, and visualizations.
- Runs up to 3000x faster on our clusters, for any size or type of images.
- Is connected to your data through our cloud integrations and beautiful Drive-like filesystem.
(6/8)
1
1
26
Of course, this fits in with the rest of the @BiodockAI platform. Use Autolabel to rapidly label your objects. Then, train an AI pipeline on those labels, which connects to your data and outputs to an intuitive results dashboard. (5/6)
1
24
Check out how to use AI Detect on our getting-started guide on docs.biodock.ai/start
Try it for free on Biodock at app.biodock.ai.
1
24
3,696
Now, with Autolabel (powered by AI), this becomes two clicks (or one drag!). Our model takes your defined box and generates what it thinks is the precise object polygon label. For most biological objects, it just works, and it’s getting more and more accurate over time. (3/6)
1
1
22
Autolabel has been trained on thousands of bio images, from fluorescent and label-free cells to H&E and colorimetric. If your objects don’t work, just let us know and we’ll see if we can add them to our training set. (4/6)
1
1
20
Try it out for free at app.biodock.ai! As always, we’ve updated our docs to provide easy step by step instructions on using SAM to power your image analysis.
2
1
20
1,881
AI Detect is built with the state-of-the-art in vision. Two large transformer models work together on multiple GPUs to get you from prompt to labels in around 2 seconds.
1
16
4,326
Tell AI Detect what to find by drawing a box around your object, which is the "prompt". AI Detect will look for, detect, and segment out all similar objects in the tile. Use the size, score, and color sliders to further optimize results. Runs in the browser, no GPU needed!
1
13
2,918
Incredibly important paper - the stark sampling graph says it all.
Phenomenal paper just published in Science persuasively demonstrating that EBV causes most if not all multiple sclerosis.
Phenomenal epidemiological design and technical achievement. 👏
science.org/doi/10.1126/scie…
2
2
11
Try it out by setting up an AI project at app.biodock.ai! Read our updated docs at docs.biodock.ai. (5/5)
10
913
Plot twist: The new name hasn't been decided yet. This is a crowd-sourced thread to generate some name ideas. 😂
1
9
It's fast! GPUs work behind the scenes work to prep tiles for SAM in ~1 second on-demand, and fast on-hover inference means you can quickly evaluate choices without even clicking. Keyboard shortcuts to finalize objects makes your workflow easy.
1
9
1,980
Try it out for free at app.biodock.ai! Or, you read more about how it works at docs.biodock.ai
2
8
1,171
You can use the labels immediately in what we call an "AI-assisted analysis," which goes directly to quantified results. Alternatively, train a refined custom model in one click for automated batch analysis.
1
7
2,270
A hidden benefit: size groups help you get better performance with less labeled data. Separate your rare cell types from your dense cell types, and you'll be able to completely separate your labeling, allowing you to label only the amount of cells needed for each group. (4/5)
1
7
999
This is one of the smartest moves YC has ever done. Great for YC. Counterintuitively, this is mostly bad for YC startups, and bad for seed stage investors.
We're excited to announce our new standard deal at Y Combinator.
When a company is accepted into YC, we now invest a total of $500,000.
@gralston shares more on our blog:
blog.ycombinator.com/ycs-sta…
1
1
7
And of course, once you’ve built a model, you now have an end-to-end model that can be run in one click on new images, on batches of any size. Results output directly to a beautiful statistics dashboard with full ability to download data. (9/10)
1
1
6
1,117
But great idea to put those metrics in the labeler too... just put that on our development timeline.
7
For more difficult images, AI Detect might not get 100% of the objects. But it’s easy to use our other AI-assisted tools to quickly grab the rest, like AI Click. We’ll be making improvements along the way - fine-tuning on more data and upgrading the models behind the magic.
1
6
2,533
Once you have a first AI model, instead of laboriously labeling every cell or tissue region in your image, you can just click big tiles (size customizable) to generate labels with AI Prelabel, which does the bulk of the work. (2/10)
1
1
6
777
But you aren’t just passively receiving labels from AI Prelabel! You’ll actively review and correct any inaccuracies. This means the model learns from you – the expert – and continuously improves. This is called human-in-the-loop learning. (3/10)
1
1
5
699
This training loop is so powerful that your first model can be trained on as little as 50 objects (cells, organoids, puncta, etc.), which takes just a few minutes using our AI-assisted labeling tools. This model will improve rapidly in one or two iterations. (6/10)
1
1
5
600
Being able to control this scale is critically important for bio images, where object sizing can vary drastically. We even have an interactive visual guide we overlay on your images to help with sizing.
1
5
1,552
It’s super easy to use AI Prelabel and the human-in-the-loop learning system, because we’ve included a step-by-step tutorial in our user-friendly model creation dashboard that walks you to a great model. No code required. (8/10)
1
1
5
666
Another important feature is the ability to order SAM’s predictions by size, depending on what you want. Change to “Smallest” to prioritize small objects like cells and organelles. Choose “Largest”, and you’ll get larger tissue regions and structures from the same click.
1
4
1,492
It’s all about efficiency. With AI Prelabel, you can retrain your model on MUCH more data than if you were labeling manually. More data = more accurate models = better quality biological research. (4/10)
1
1
4
668
Public API documentation at docs.biodock.ai/public-api-b…. All users can generate their API key in-platform at app.biodock.ai. Biodock and our API are free for academics. Our API is free while in beta for enterprise users. (5/6)
1
5
182
There’s also a great hidden benefits of AI Prelabel: fast adaptation to novel images. If your images look very different and your performance degrades, it’s simple to fix. Just AI Prelabel, correct the mistakes, and retrain. (7/10)
1
1
4
580
The key feature on Biodock is the ability to adjust the scale that SAM “sees” your images. Choose a smaller tile size to help SAM detect small objects like cells. Choose a larger size for organoids or tissue regions.
1
4
1,654
It also means that any work you do as part of the system is not used on areas where the AI model is already doing well, but focused on areas where your AI model tripped up, meaning that you are getting the best value from your time. (5/10)
1
1
4
629
The API works seamlessly with our GUI. Every file upload and analysis job invoked through the API is also saved and accessible in our easy-to-navigate dashboard. Start an analysis from a POST request and view results in your browser! (4/6)
1
4
163
Batteries included: serverless infrastructure spins up on-demand to handle even large jobs in 3-5 minutes. Progress built into all long-running jobs. Big images are seamlessly tiled and stitched during analysis. (3/6)
1
4
173
This is only (very!) good for YC startups that struggle to raise. However, the MFN clause will hurt oversubscribed rounds, disincentivize small, lower priced checks from value investors, and push valuations up. Again, smart for YC.
1
3
Indirectly, yes. If you label and train on your objects of interest, then your new AI pipeline will output area, X-Y position, length, width, solidity, and more!
1
2
Trained from scratch on an image transformer backbone. Lots of ground to cover here in transferring more effectively from our bio data, and few-shot learning to make labeling iterative and easier!
3
Yes! You can easily create classes for each type you are looking at, and then label them using Autolabel. When your pipeline is trained, you’ll have an end to end analysis solution that can count class and morphology.
1
2
Super easy! Hit 'e' on your keyboard and you can add or subtract from labels with our built in pen tool. Finetuning the model for your dataset is available for paid engagements for now.
3
In general, we don't yet support volumetric analysis. The closest we have for now is plane-by-plane analysis. It's definitely on our radar!
1
3
With Size Groups, your AI model can zoom close to capture tiny puncta, zoom out a little to segment cells, and zoom out still further for regions. This increases accuracy and reduces stitching artifacts because each object gets the right amount of neighboring context. (2/5)
1
3
111
Easily build: a script that segments batches of cell culture images every 4 hours, a data sync program for multiple imagers, an alarm that runs QC on large tissue images, or an internal dashboard for image analysis submission. (2/6)
1
3
201
Size groups also unlock interesting insights between large and small objects! Set up a hierarchy in Biodock to automatically count the number of cells per intestinal crypt, the number of marker positive cells per brain region, or get a cell count per organoid. 🔢 (3/5)
1
3
104
Hey John! Ilastik is awesome - I think there's a few key differences:
1. Biodock focuses mostly on deep, image transformer based models which are more powerful (also take more GPU juice to train)
1
3
3. A couple of other things: we're in the cloud, connect to your data, provide out of the box smooth image viewing, and have some nice UX for labeling and throughout the platform!
1
3
Hi Aftab! You can't see metrics in the labeler, but once you train an AI pipeline, you get many metrics. One of them is average integrated intensity across every channel in the image for your ROI!
1
Great idea! Not all of our users are medical image users, but I think it would be a great option for those that are 😀
2
122
Biodock might be able to help here! We'd love your feedback @realjakewinter @RutterLab @BiodockAI
2
Yes! We have Pearson and Spearman colocalization currently and support up to 100 channels. Train a model to find your cells, and then add colocalization metrics to your model.
1
2
173
2. Training a model with Biodock also automatically gives you access to a powerful filesystem, run infrastructure, and dashboard for any size of batch
1
2
You could certainly do this with Biodock:
1. Create classes for each of the stages you want to identify.
2. Use Autolabel to label these classes quickly.
3. Train and get an end-to-end pipeline to identify the classes and morphological information, counts, etc.
1
1
Awesome! Shoot me a message here on Twitter or michael@biodock.ai if you run into any issues or questions.
2
120
Love to hear it! We'll be shipping more improvements to these models so that we can perform well even in difficult use cases.
1
113
It should! Choose a small tile size. If you have spots that are stuck together you might need to use our AI Select tool to make a few corrections.
1
1
865
You use SAM + our AI annotation tools to build your model, in about 15-20 minutes of work. Then you can use that model to batch process images!
1
1
27
The segmentation part is pretty easy in this case. The more difficult part is precise one-shot object detection off a prompt, which is why we think this is pretty cool!
342
I'm personally really excited about the space - I think there's an inflection point developing. Especially interested in improving models over time and low-data training in vision and beyond. All in easy to use interfaces.
1
So do we... incredibly painful, doesn't scale, soul-sucking. Hopefully the next generation of scientists doesn't need to do manual circling ever again!
1
96
For some it could be, but we have lots of academic scientists who do it! You can also connect an S3 bucket, and we're SOC 2 Type II certified - the gold standard for data security.
1
78