jpcbertoldo/gfpa66-volumes

Glass fiber-reinforced polyamide 66 3D X-ray computed tomography (XCT) dataset for deep learning segmentation

todos

• updated zip files links
• updated h5/xdmf files links
• updated pymicro tutorial link
• make the read_h5 work
• update description in zenodo
• include link to this
• todo show how to do it with fiji --> add it to zenodo as well

Please cite us as:

bibtex entry TO BE UPDATED

This repository is a complement to paper above. We show how to read the data mentioned in the paper and how to load the Keras models we trained to segement that data. Both the data and the models are publicly available on Zenodo.

Links:

• data:
• models:

data

We provide annotated XCT volumes (stack of 2D images) of two specimen of PolyAmide 66 (PA66) reinforced with glass fibers.

There are specimen:

• pa66 or train-val-test: for wich we provide the gray-level image, its manual segmentation (voxel-wise annotation volume) with 3 classes, and the segementation predicted by one our models on the test partition of the volume.
• crack: another speciment of the same material, segmented with the same model.

We show how to open data in three ways:

• from .raw files in python

• from .raw files using Fiji (ImageJ)

• from .h5 files in python using pymicro

files

.zip file	.raw file	Description
pa66.zip	pa66.raw	Data (gray level image stack) of the Train-Val-Test volume (1900 slices).
	pa66.ground_truth.raw	Ground truth segmentation of the Train-Val-Test volume.
pa66_test.zip	pa66.test.prediction.raw	Segmentation generated by the best 2D model on the test set (last 300 slices of the Train-Val-Test volume).
	pa66.test.error_volume.raw	Disagreement between the ground truth and the model's prediction on the test set: 1 means incorrect, 0 means correct
crack.zip	crack.raw	Data of the non-annotated volume containing a crack inside.
	crack.prediction.raw	Segmentation generated with the best 2D model on the crack volume.

Each .raw file has a corresponding .raw.info file, containing metadata (data type and volume dimensions).

The .h5 contains all the volumes mentioned above. It also has a .xdmf metadata file.

You will find all these files here:

.raw with python

**There is no special dependency to use it this way (apart from basic stuff: numpy and matplotlib if you want to plot). ** 51

.raw with Fiji (ImageJ)

.h5 with python

**There is no special dependency to use it this way (apart from basic stuff: numpy and matplotlib if you want to plot). **

You only need the .zips and the module read_raw.py in this directory.

Download and extract the contents of the crack.zip, pa66.zip, and pa66_test.zip files to this directory.

TODO add links

The .info files contain the shape (x, y, and z sizes) of the corresponding files with the same name so that you don't need to type them.

The module read_raw.py has a function that will be used to load the files' contents into numpy.ndarrays - credits to pymicro.

See the notebook read_raw.ipynb for a brief demo of how to open the .raws as numpy arrays - put the contents of the .zip files in a folder raws before.

SampleData (.h5 + .xdmf)

Download the files below and put them in this directory.

• pa66_volumes.h5
• pa66_volumes.xdmf

Install pymicro and follow the tutorial in read_h5.ipynb.

To see how the .h5/.xdmf files were generated, see the file write_h5.ipynb or check this tutorial in pymicro's repository'.

models

Glass fiber-reinforced polyamide 66 3D X-ray computed tomography dataset for deep learning segmentation by Joao P C Bertoldo, Etienne Decencière, David Ryckelynck, and Henry Proudhon is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.