DeepACSON

Automated Segmentation of White Matter in 3D Electron Microscopy

Flowchart of the DeepACSON segmentation of myelinated axons:

DeepACSON was developed by Ali Abdollahzadeh at the University of Eastern Finland to trace the entirety of myelinated axons in low-resolution, large field-of-view, 3D electron microscopy images of white matter.

A. Abdollahzadeh, I. Belevich, E. Jokitalo, A. Sierra, J. Tohka, DeepACSON: Automated Segmentation of White Matter in 3D Electron Microscopy, bioRxivdoi:https://doi.org/10.1101/828541.

A. Abdollahzadeh, A. Sierra, J. Tohka, Cylindrical shape decomposition for 3D segmentation of tubular objects, arXiv:1911.00571v2 [cs.CV] (2019). URL http://arxiv.org/abs/1911.00571.

BM4D denoising

We used the BM4D filter to denoise 3D-electron microscopy images: one can download BM4D v3.2 from https://www.cs.tut.fi/~foi/GCF-BM3D/. The BM4D algorithm can be run in its low-complexity mode to reduce the computation time. To conform to RAM requirements or enabling parallel denoising on CPU clusters divide the big 3D-EM volumes into non-overlapping patches.

Training and inference

Install Elktronn as instructed in https://github.com/ELEKTRONN. The network can be trained with the train.py script, which expects .h5 files as training materials. Use the inference.py script for the semantic segmentation of ultrastructures.

We provided a pre-trained network in ./models/pretrained_mAxon.mdl, as in ./train.py, to be applied to BM4D denoised electron microscopy images. The network was trained on 3D-EM volumes of 50 nm x 50 nm x 50 nm resolution. The semantic segmentation includes myelin, intra-axonal space of myelinated axons, and mitochondria.

If you want to skip installing ELEKTRONN, we included a 3D U-Net with residual blocks, written for PyTorch, in ./models. Please, install pytorch>=1.5.0.

Instance segmentation

The cylindrical shape decomposition (CSD) algorithm in DeepACSON is currently supported for Python 2.

Installation

The CSD algorithm requires no installation. To install the necessary dependencies, run:

pip install -r requirements.txt

Due to the size of datasets, we provided examples as 3d-coordinates and their corresponding bounding box in dict files: 'objSz', 'objVoxInx'.

The mAxon_mError file is an example of an under-segmentation error, where two myelinated axons incorrectly were merged. The bounding box of this 3D object is 2004x1243x180, acquired at 50 nm x 50 nm x 50 nm resolution. To retrieve an object, load the coordinates as:

import numpy as np

fn = './example/mAxon_mError.npz'
data = np.load(fn)

obj_sz = data['objSz']
obj_voxInd = data['objVoxInx']
bw = np.zeros(obj_sz, dtype=np.uint8)
for ii in obj_voxInd: bw[tuple(ii)]=1

Skeletonization

To skeletonize a 3D voxel-based object, we implemented the method from Hassouna & Farag (CVPR 2005); the method detects ridges in the distance field of the object surface. If you use skeleton3D in your research, please cite:

M. Hassouna and A. Farag, Robust centerline extraction framework using level sets, DOI 10.1109/CVPR.2005.306
A. Abdollahzadeh, A. Sierra, J. Tohka, Cylindrical shape decomposition for 3D segmentation of tubular objects, arXiv:1911.00571v2 [cs.CV] (2019). URL http://arxiv.org/abs/1911.00571.

The implementation only requires numpy and scikit-fmm

from skeleton3D import skeleton
skel = skeleton(bw)

You can modify the code to define the shortest path either as

Euler shortest path: sub-voxel precise
discrete shortest path: more robust but voxel precise

and also to exclude branches shorter than length_threshold value.

With the default values, one should be able to reproduce the results in ./results/skeleton_mAx.npy. The skeletonization of the example object, 2004x1243x180 voxels, takes about 9 mins on a machine with 2xIntel Xeon E5 2630 CPU 2.4 GHz with 512 GB RAM.

Cylindrical shape decomposition

To apply CSD on a 3D voxel-based object, given its skeleton as skel, we have:

from shape_decomposition import object_analysis
decomposed_image, decomposed_skeleton = object_analysis(bw, skel)

decomposed_image is a list, which length is equal to the number of decomposition components. The bounding box for each component is equal to the bounding box of BW image, the input object. In ./results, we provided dec_obj1 and dec_obj2 for the mAxon_mError file with an under-segmentation error. Due to the size of datasets, we provided the decomposed objects as 3d-coordinates and their corresponding bounding box in dict files: 'objSz', 'objVoxInx'.

import numpy as np

fn = './example/dec_obj1.npz'
data = np.load(fn)

obj_sz = data['objSz']
obj_voxInd = data['objVoxInx']
obj1 = np.zeros(obj_sz, dtype=np.uint8)
for ii in obj_voxInd: obj1[tuple(ii)]=1

The same holds for obj2.

You can modify the code to define H_th value, which lies in the range [0, 1]. H_th is the similarity threshold between cross-sectional contours. Also, setting ub and lb, where ub>lb also modifies the zone of interest.

DeepACSON instance segmentation, i.e., the CSD algorithm, can be run for every 3D object, i.e., an axon, independently on a CPU core. Therefore, instance segmentation of N axons can be run in parallel, where N is the number of CPU cores assigned for the segmentation task.

aAbdz / DeepACSON