__Address of ATLaSUI (BigBrain1 pre-selected):__[link](http://medpc076.ime.kfa-juelich.de:5001/microdraw/microdraw.html?source=data/brainmap/B20.json)
## What is ATLaSUI?
*ATLaSUI* (Automatic Tissue Labeling System UI) is a web application that allows users to automatically create segmentations of cytoarchitectonic areas using a deep learning algorithm.
It allows the import of existing annotations or the creation of new annotations, and feed these annotations as training data to multiple deep neural network models.
If all goes well, the trained neural network models are able to automatically identify the respective brain area in other sections, essentially "filling the gaps" between manually created annotations.
This workflow thus allows to map brain areas across large series of sections, which can serve as basis for 3D reconstruction or dense analysis.
## How does it work? - An overview
The following points detail a typical workflow of using ATLaSUI.
The points intent to give a high-level overview of the steps, which are further detailed in the remainder of this documentation.
1. Create a user account for the web page.
2. Import annotations into the user account. Annotations can be imported from [Online Section Tracer (OST)](https://jubrain.fz-juelich.de/apps/sectiontracer/sectiontracer.php) or from [Microdraw](http://ime262.ime.kfa-juelich.de:3000/data?source=http://ime262.ime.kfa-juelich.de/microdraw/data/brainmap/B20.json). Alternatively, annotations can be created directly in ATLaSUI.
3. Define project and tasks. Each task correspond to one trained deep neural network model, and holds annotations to be used for training. Projects group tasks that are related to each other (e.g., regarding similar brain regions).
4. Run training and prediction jobs to train neural network models on the created tasks.
5. Inspect the prediction results and note down the quality of the created predictions per section.
6. (optional) If the results are not satisfactory, repeat steps 2-5 by providing more annotations and training more neural network models.
7. (optional, currently BigBrain1 only) Send the results of your quality control to Christian Schiffer to attempt a 3D reconstruction of the segmented areas.
## Pre-requisites
The following pre-requisites have to be met to enable use of ATLaSUI:
- __FZJ VPN access__ ATLaSUI displays sensible data and accesses restricted compute infrastructure. For this reason, ATLaSUI is currently only accessible while connected to the VPN of Research Centre Jülich, or from within the JuNet. If you don't have VPN access already, please coordinate with your contact person at FZJ to learn how you can apply for VPN access. Note that external mail addresses require special approval be the IT services. Contact the INM-1 secretaries to get the approval.
- __ATLaSUI user account__ All data used and created by ATLaSUI is associated with a user account. You can simply create a user account yourself by clicking `Log in` in the top left corner of the page, then click `Register` at the bottom of the popup window. Enter a username, your mail, and select a password. The username should reflect who you are, and can include information about the project you are working on (e.g., `cschiffer_visual_system`). Note down your username, as you will need it to log in and for importing annotations. __Note:__ Please note down your password safely, as there is currently no way to recover lost passwords!
- __Annotations__ ATLaSUI is integrated into a modified version of the visualization and annotation tool *Microdraw*, so it is possible to create annotations directly in ATLaSUI. However, it is more common to already have annotations available in a different tool, which are then imported into ATLaSUI. The following import options are currently available:
- __Importing from standard Microdraw__ Annotations created in the [standard Microdraw](http://ime262.ime.kfa-juelich.de:3000/data?source=http://ime262.ime.kfa-juelich.de/microdraw/data/brainmap/B20.json) of the INM-1 can be imported. Please contact Christian Schiffer to trigger the import. __Specify the username to import annotations from, and the username to import annotations to!__
- __Importing from Online Section Tracer (OST)__ Annotations created in the [Online Section Tracer (OST)](https://jubrain.fz-juelich.de/apps/sectiontracer/sectiontracer.php) can be imported in a similar fashion. Please contact Christian Schiffer to trigger the import. __Specify the OST project to import annotations from, and the username to import annotations to!__ Data from multiple projects can be imported into the same ATLaSUI account. __Note:__ Since OST handles image resolution slightly different than ATLaSUI and Microdraw. As a result, annotations for some brain sections cannot be imported into ATLaSUI. If applicable for a project, these annotations have to be re-created in ATLaSUI using the provided annotation tools.
- __Importing from Offline Section Tracer XML files (deprecated)__ It is possible to import XML files exported form offline section tracer. To do so, send the XML files to Christian Schiffer, and specify the username to import the annotations to.
- __Annotating in ATLaSUI__ When annotating in ATLaSUI directly, please stick to the naming scheme known from OST: `NAME_l_0`, where `l` can be `l` or `r` and `0` can be a number. Avoid adding additional parts to the name. It is important to know that ATLaSUI identifies areas by their name (capitalization does not matter). Areas with different names are considered as different areas, leading to incorrect results!
## Core concepts
The segmentation method implemented in ATLaSUI follows a relatively simple idea:
A user annotates one or several brain regions in a set of sections.
These annotations are used to train neural networks, which then segment all sections in-between the annotated sections.
Although this principle is identical for all use cases, there are some questions that have to be answered depending on the area:
### How to subdivide the section range?
Consider we want to segment an area (or multiple) in sections 1-1000, where sections 1,100,200,...,900,1000 are annotated.
We can
1. use all sections to train one model, so the model is trained on 1,100,200,...,900,1000, or
2. we can subdivide the range into (usually equally sized) intervals. In this case, we train one model on 1 & 100, a second model on 100 & 200, ..., and the final model on 900 & 1000.
The first method is suggested if the structure is small and/or has a relatively regular shape and structure across the entire extent.
For example, nuclei with a round or almond-shaped structure have turned out to be nicely segmented using a single model.
In comparison, areas that vary a lot in microstructure and/or morphology (e.g., most cortical areas) are typically better segmented using the interval approach.
Note that the above assumes equally sized intervals, but this is not required.
For example, it has turned out to be beneficial in many cases to use wide intervals in the center of an area (along the cutting direction), but a finer interval towards the boundaries.
If in doubt, it is recommended to start with the simpler approach (one model for all), and increase the complexity (using intervals) if necessary.
__Tip:__ The task wizard supports users in automatically creating tasks at certain intervals.
### How to subdivide multiple areas?
Consider you are interested in brain areas A, B, and C.
Similar to the section subdivision, you have multiple options to handle the areas:
1. Train one model to segment areas A, B, and C, or
2. Train one model for each area, meaning one model for A, one model for B, and one model for C.
Again, it is advisable to start with the simple approach of training one model for all considered areas, and go for the more complex approach if the first approach does not yield good results.
Note that the subdivision of sections and areas can be freely combined, e.g.,
1. Training one model for all sections and all areas.
2. Train one model for each section interval and all areas.
3. Training models for all sections, but one model per areas.
4. Train one model per section interval and area.
The above cases increase in complexity.
It is usually a good approach to start with the simplest (top) and progress towards the more complex ones (bottom) only when necessary.
Note that mixing approaches is always possible, as described in the previous section about intervals with varying depth.
*Example use case for training deep learning models. This example shows the last case indicated above: Models are trained for each section interval, and for one specific area.*
## FAQ
### Q: Where can I find ATLaSUI?
A: ATLaSUI is currently hosted [here](http://medpc076.ime.kfa-juelich.de:5001/microdraw/microdraw.html?source=data/brainmap/B20.json).
Note that this URL shows the BigBrain dataset by default.
If you want to use different datasets, please contact Christian Schiffer to receive an appropriate URL.
### Q: Who is responsible for ATLaSUI?
A: ATLaSUI is developed, maintained, and supported by Christian Schiffer (c.schiffer@fz-juelich.de).
