12 Jan 2021

FIBSEM Mitochondria Segmentation

Pre-trained model for the Deep Learning Prediction module detecting mitochondria in a Focused Ion Beam Scanning Electron Microscope (FIBSEM) data set

This is a pre-trained model for detecting mitochondria in a Focused Ion Beam  Scanning Electron Microscope (FIBSEM) data set.

It is most useful to detect bright mitochondria on a darker background.

The model is based on a convolutional neural network (CNN) with U-Net architecture and was trained on a FIBSEM data from human liver tissue. 

The data was pre-processed with the BSE SEM denoiser model: (https://xtras.amira-avizo.com/xtras/bse-sem-denoiser).



The model can be tested with the FIBSEM_Human_Liver_sample.tif sample data


The downloadable archive contains the following files:
•    Mitos_Adam_final.json = model architecture,
•    Mitos_Adam_final.hdf5 = trained weights,
•    Mitos_Adam_final.py    = pre-processing functions required to connect the model to Avizo2D / Amira / Avizo grayscale images.
•    a test image (FIBSEM_Human_Liver_sample.tif)


The model can be applied to any grayscale image or volume, using the module Deep Learning Prediction.
Please note that the training data was pre-processed with the BSE SEM denoiser model (https://xtras.amira-avizo.com/xtras/bse-sem-denoiser). This is recommended before using this model to predict mitochondria.

Please follow the example video to see the model applied in Avizo2D to the sample data set. Amira and Avizo users can follow similar steps using the Image Stack Processing module and workroom.

FIBSEM_Mitochondria_Segmentation_zoom

Data courtesy of Bruno Humbel, Caroline Kizilyaprak, and Jean Daraspe (Electron Microscopy Facility Lausanne University)