neurofly

This package provides tools for semi-automatic neuron reconstruction. Features based on deep learning, like image segmentation and deconvolution are implemented in tinygrad, which can run on almost any GPU (NVIDIA, AMD, Apple, Qualcomm, Intel).

Installation¶

Install the latest version

pip install --upgrade git+https://github.com/beanli161514/neurofly.git

or

pip install neurofly

Dataset¶

We provide several expert-proofread reconstruction results for testing, model training, and evaluation. Zenodo Link

Content of samples¶

name	size	species	label type	imaging
rm009_labeled	629MB	macaque	skeleton	VISoR
mouse_labeled	260MB	mouse	skeleton	VISoR
z002_labeled	204MB	mouse	skeleton	VISoR
fmost_labeled	370MB	mouse	skeleton	fMOST
RM009_noisy_1	65MB	macaque	morphology	VISoR
RM009_noisy_2	65MB	macaque	morphology	VISoR
fmost_test	65MB	mouse	morphology	fMOST
z002_dendrites	768MB	mouse	morphology	VISoR
RM009_arbor_1	288MB	macaque	morphology	VISoR
RM009_axons_1	600MB	macaque	morphology	VISoR
RM009_axons_2	600MB	macaque	morphology	VISoR
z002	8.92G	mouse	morphology*	VISoR

$*$ annotation in progress

Label format¶

Morphology labels are graphs saved in SQLite database with 3 tables:

segments	nodes	edges
sid	nid	src
points	coord	des
sampled_points	creator	date
	status	creator
	type
	date
	checked

Segments are results of the segmentation stage, they are used to generate initial nodes and edges.

Basic usage example¶

NeuroFly packaged 4 napari plugins for image browsing, image segmentation, and data annotation.

Segmentation¶

NeuroFly supports whole brain image saved in hierarchical data structures(ims, h5, and zarr) in Imaris File Format and small image volumes saved in single-channel tiff format. Here we use a mouse brain in our dataset named z002.zarr.zip as example.

This brain is sparsely labeled, which means only a tiny puny part of neurons are lighted and imaged. To extract these foreground singals, you can use the provided command line interface 'seg'. We provide a default weight trained on images captured by VISoR and fMOST.

seg -i z002.zarr.zip -vis -d z002.db

or use the graphical interface

This process may take about 10 hours depending on your you hardware configuration. When finished, you should see the extracted segments and a database file named z002.db in your working dictionary.

An image block with severe contamination and the segmentation result

Manual connection and proofreading¶

Launch annotation tool from napari menu, Plugin -> neurofly -> Segs Annotator

Load data¶

Load image file (z002.zarr.zip) and database file (z002.db), then click refresh panorama button to show the panorama view.

Select one node as start point of annotation¶

In panorama mode, you can easily identify sparse, bright signals that are promising for reconstruction. The silde bars 'short segs filter', 'length thres', and 'point size' can be adjusted to hide noise and short segments.

If you can clearly identify foreground segments, click on one of the conspicuous segments to select it as start point of annotation. Once selected, the id of picked node will be displayed at node selection. Then click 'switch mode' to switch to labeling mode, and the tasks will be generated automatically.

Task generation¶

Given a selected node, task generator analyses its connected component and extract all unchecked terminal nodes. The tasks are designed very simple: Connect the center node with the surrounding nodes if there should be an edge. The criterion is whether the edge aligns well with the imaged neuron fibers.

Node operations¶

In each task, a center node and nearby segments are rendered, you can add/remove nodes and edges to get a reasonable local structure.

Left click on nodes to add/remove an edge between it and the center node

Right click to remove a node

Press 'g' or use left panel to switch to 'image' layer, then right click to add points

Use dropdown selection in right panel to add type label for center node.

Deconvolution¶

Press 'i' or click on 'deconvolution' to deconvolve the image

Proofreading¶

If you find something wrong when labeling, for example, two somas are connected together. You can use proofreading mode to check the neuron branch by branch.

Performance¶

NeuroFly supports rendering of more than ten million points. (tested on M3 Macbook Air and RTX 3090 workstation)

Export as swc file¶

Switch to panorama mode, adjust 'length_thres' to filter out short segments and keep only complete neurons. Then press 'export swc files', each neuron will be saved as one .swc file in your working dictionary.