Skip to content

Segment Properties

Segment properties are a method of organizing information about segment IDs in the current Google/Spelunker version of Neuroglancer. Each segmentation layer can have at most one set of segment properties, expressed as an additional source for the layer. The segment properties themselves are simply a JSON file following a particular specification, that allows you to associate labels, descriptions, tags, and numerical values with each segment ID.

Basics

Segment properties consist of a list of segment IDs and a collection of properties for each ID in the list. There are four types of properties that can be viewed in Neuroglancer:

  • Label: A string that names each segment ID. Only one per ID per property list.
  • Description: A longer description of each segment ID. Only one per ID per property list.
  • Tags: A small collection of strings that can be used to filter segments. There can be more than one tag per ID.
  • Number: Numeric properties (e.g. depth, synapse count) for each cell. There can be more than one numeric property provided.

In addition, the specification allows string properties, but these are not currently shown in Neuroglancer.

Segment property example
Segment properties help make Neuroglancer more browsable from data

Building segment properties

The easiest way to work with segment properties is to build them from a dataframe. The only absolute requirement is that the dataframe has a column associated with segment IDs. For example, assume we have a pandas dataframe df with columns segment_id, soma_depth, and cell_type with some data in them. We can use the SegmentProperties.from_dataframe method to build a SegmentProperties object from this dataframe. In this case, we want soma_depth to be a numeric property and the values cell_type to be a tag.

import pandas as pd
from nglui.segmentprops import SegmentProperties

df = pd.DataFrame(
    {'segment_id': [1, 2, 3],
     'soma_depth': [100, 200, 300],
     'cell_type': ['pyramidal_cell', 'interneuron', 'interneuron']}
)

seg_prop = SegmentProperties.from_dataframe(
    df,
    id_col='segment_id',
    number_cols='soma_depth',
    tag_value_cols='cell_type'
)

To get a formatted segment proerties, you can use the to_dict method:

seg_prop.to_dict()

which will produce something like:

{'@type': 'neuroglancer_segment_properties',
 'inline': {'ids': ['1', '2', '3'],
  'properties': [{'id': 'tags',
    'type': 'tags',
    'tags': ['interneuron', 'pyramidal_cell'],
    'values': [[1], [0], [0]]},
   {'id': 'soma_depth',
    'type': 'number',
    'values': [100, 200, 300],
    'data_type': 'int32'}]}}

Note that multiple numeric properties can be passed a list of column names.

Tag properties

There are two ways to associate tags with segment IDs. When using the tag_value_cols argument, the unique values (except None) across all specified columns are used as tags. Values that are duplicated across different columns will be counted as a single tag. For tag value columns, the order is determined by the sorting of values. By default, this is alphabetical, but it will follow the order from categorical data in pandas.

If you need more fine control over the tags, you can use the tag_bool_cols argument. Here, a column name or list of column names are passed with the expectation that each column name represents a tag and if the value of a given row is True, the tag is associated with the segment ID. Here, the order will follow the order of the list.

Both tag_value_cols and tag_bool_cols can be used together, with the order being values followed by boolean columns.

You can also specify tag_descriptions, a dictionary where the tag names are the keys and longer-form descriptions are the values. Any tag without a description will be passed through directly.

Number properties

If you look at the JSON output above, you see that the soma_depth property has a field "data_type" that specifies the type of the numerical property, in this case int32. Data type values are inferred from the column dtype and safe conversions are validated.

Random number columns

Often, segment properties are used for very large lists of root ids. If you want to sample just a few percent of these in Neuroglancer, it can be hard to get a small number of random ids without a lot of clicking. Adding a numerical property that is just a random number between 0 and 1 will give you a way to load small subsets of the data using the filter tools. You can add random columns when building a SegmentProperty object from a dataframe by setting the random_columns argument with the number of extra columns you want to add. For example, to add a single random column:

seg_prop = SegmentProperties.from_dataframe(
    df,
    id_col='segment_id',
    number_cols='soma_depth',
    random_columns=1,
)

Will add one random column as a numerical property with the default name random_sample. Using numbers higher than 1 will add multiple random columns with names random_sample_0, random_sample_1, etc. The name of the random column can be changed with the random_column_prefix argument. Note that a value error will be raised if you try to add a random column with a name that already exists in the dataframe.

Building segment properties manually

It is also possible to build segment properties manually with the SegmentProperties class. Each type of property is represented by a different class: LabelProperty, DescriptionProperty, TagProperty, and NumberProperty. Each property type has required and optional arguments that are used to build the property. More details can be found in the API reference. You can then either build a SegmentProperties object from these properties, for example:

from nglui.segmentprops import SegmentProperties, LabelProperty, NumberProperty

labels = LabelProperty(values=['cell_1', 'cell_2', 'cell_3'])
numbers = NumberProperty(
    id='hugeness',
    values=[1,2,3],
    data_type='float32',
)
props = SegmentProperties(
    ids = [1,2,3],
    label_property=labels,
    number_properties=numbers, # Could be a list of NumberProperty objects as well.
)

Note that validation happens on the to_dict method, so you can build invalid properties (such as different-length lists of ids and property values) without error.

Saving segment properties

Segment properties must be hosted on a server to be used in Neuroglancer. There are two main approaches to this:

1. Use a CAVE state server endpoint to upload properties.

The CAVE state server has a special endpoint you can use to upload segment properties. After initializing a CAVEclient object, you first upload the properties with the upload_property_json method and then build a link from the resulting id you get back.

from caveclient import CAVEclient
client = CAVEclient('MY_DATASTACK')
ngl_url = "https://neuroglancer-demo.appspot.com"

prop_id = client.state.upload_property_json(seg_prop.to_dict())
prop_url = client.state.build_neuroglancer_url(prop_id, ngl_url=ngl_url, format_properties=True)

You will then use the resulting URL in the source field of a segmentation layer in Neuroglancer.

This approach is convenient, but there is a limit to how large the properties list can be. In addition, you do not get to choose a readable URL for the properties and they must be behind CAVE authentication.

2. Save the properties to a file web-accessible location.

For large properties or properties you want to be publicly accessible, you can save the properties to a file and host them on a web-accessible location with a particular organization.

First, save the properties to a file called info (no extensions!)

import json

with open('info', 'w') as f:
    json.dump(props.to_dict(), f)

Next, create a directory in a web-accessible location (e.g. https://my-server/my-property) or public cloud bucket (e.g. gs://my-bucket/my-property) and upload the info file to that directory.

The URL you will use to load the properties into Neuroglancer will be precomputed:// followed by the URL of the directory or bucket (e.g. precomputed://https://my-server/my-property or precomputed://gs://my-bucket/my-property).

Adding segment properties to Neuroglancer

If saved segment properties with either of the two methods above, you will get a URL that you can load into Neuroglancer as a source for a segmentation layer. If you want to do this manually in a browser you have already have open, select the segmentation layer, go to the source tab, and click the + button to add a new source and paste the URL in.

Segment properties in StateBuilder.

The StateBuilder class lets you set either static segment properties or segment properties that are generated from a dataframe. See the StateBuilder documentation for more details.