Volume Segmentation Tool with GUI

This repository holds the script for Volume Segmentation Tool and its Gradio-based Webui.

Paper coming soon...

Supporting Operating Systems includes: Windows and Linux and macOS.

About the Tool

Volume Segmentation Tool is a python based tool that utilizes deep learning to perform volumetric electron microscopy image segmentations.

Based on Pytorch backend. Includes a Gradio based graphical user interface.

The desired images are 8bit or 16bit grey scale 3D tif/hdf as commonly generated by electron microscopes.

This tool does not support 2D image segmentation, nor colorful image segmentation. In which case you should consider ilastik or Trainable Weka Segmentation.

Advantages

• High efficiency
  • Custom implemented augmentation specialised in volumetric images
  • Process many depth slices all once, rather than slice by slice
  • Support fp16/bf16 training
  • Up to 20x speedup compare to nnUNet
• Versatile
  • Semantic Segmentation & Instance Segmentation
  • Isotropic & anisotropic images
  • Adaptive network type, size and depth
  • Support AMD GPU (only under Linux)
• Easy to use
  • One click installation script
  • Graphical User Interface
  • Visualisation tools for data augmentation and network activations
  • Visualise network output on the fly
  • TensorBoard Logging & Can export as Excel spreadsheet

Limitations

• Does not support 2D images, nor images with colours (grey scale only)
• Since it's based on Deep Learning, the tool needs to be used with a discrete GPU
  • It's possible to run this tool without a discrete GPU, however it will run extremely slowly
  • Recommended minimal GPU requirement: 4GB of Video Memory, made by Nvidia or AMD
• Since it's based on Deep Learning, the user has to create training samples to train the network before it could perform segmentation
  • For more information, please see Tutorials.

Installation

1. Install Python 3.10.

   • During the installation process, ensure that you select the option to add Python to the 'PATH' environment variable.

2. Install Git.

   • If you are using Linux, it's highly likely you can skip this step.

3. Open a terminal and navigate to the desired installation directory.

4. Clone the repository by running the following command:

   git clone https://github.com/fgdfgfthgr-fox/Volume_Seg_Tool.git

5. Run the corresponding "install_dependencies" script for your system.

   • For Windows, it would be install_dependencies_Windows.bat

6. Wait for the script to finish, which could take a while depends on your network speed.

Starting GUI

1. Run the corresponding "start_WebUI" script for your system. This would open up a terminal.
   • For Windows, it would be start_WebUI_Windows.bat
2. Your web browser should automatically open up a "website" with url "127.0.0.1:7860" or something similar.

Tutorials

Please see the Wiki (Mostly WIP).

Evaluation

All number below were obtained with semantic segmentation

Dataset	Network Configuration	Dice Score	Peak VRAM Use
UroCell Mitochondria 5 fold cross-validate	Residual UNet, base channel of 16 and depth of 5. Patch size 128x128x128.	0.9277	6870MB
Lucchi++ Mitochondria 2 fold cross-validate	Residual UNet, base channel of 16 and depth of 5. Patch size 256x128x128.	0.9193	10200MB
ISBI 2012	Residual UNet, base channel of 16 and depth of 5. Patch size 256x256x16.	0.7868 V_Rand:0.9725 V_Info:0.9852	1812MB
Kasthuri++ 2 fold cross-validate	Residual UNet, base channel of 16 and depth of 5. Patch size 384x382x32.	0.9596	7958MB

Ground Truth Segmentation of one of the UroCell Dataset sub-volumes showing the mitochondria:

Model prediction using information learned from the 4 other sub-volumes, Dice score of 0.9238:

(Image rendered using MIB)

Credits

This tool was developed under the scholarship funding from AgResearch, and was helped by the members of the Bostina Lab of the University of Otago. As well as Lech Szymanski from the Computer Science department.

The example dataset included in this repository was collected by Vincent Casser.

The UroCell Dataset was heavily used during the development of the tool.

Special thanks to YunBo Wang from Xidian University, who gave me exceptional helps at the early stage of development.