Robust Ensemble Approach to Automatic Segmentation of Mitochondria from FIB-SEM Images

In this repository, we have employed the Mask R-CNN model to detect and segment the mitochondrial region of interest from the UroCell dataset. We utilized the maskrcnn_resnet50_fpn backbone for building the Mask R-CNN network, but users can easily switch to another pre-trained network if desired.

Link of the datasets used in this study:

• https://www.epfl.ch/labs/cvlab/data/data-em/ (Lucchi (dataset-1))
• https://sites.google.com/view/connectomics (Lucchi++ (dataset-2))
• https://sites.google.com/view/connectomics (Kasthuri++ (dataset-3))
• https://github.com/MancaZerovnikMekuc/UroCell (UroCell (dataset-4))

Training and Testing

To train your own model with custom data, please use train_file.py. This file also provides a convenient way to test your data. You can download the UroCell dataset from (https://github.com/MancaZerovnikMekuc/UroCell).

For our 3D segmentation work, we employed the PyTorch 3D U-Net model available at (https://github.com/wolny/pytorch-3dunet). We've provided a config YAML file for both training and testing. Before using it, ensure that you've installed the pytorch-3dunet package on your workstation.

• The ensemble_fusion.py file can be used to fuse the prediction output of two models.

Results with MASK R-CNN:

Test Image 1	Valid Pred Image

Results with 3D U-Net:

3D Unet

Results with ensemble fusion method:

Mask R-CNN Output	3D U-Net Output	Proposed Ensemble Fusion Method