Interactive U-Net

A segmentation tool that utilizes the U-Net deep learning architecture to quickly and efficiently segment 3D volumetric images. It utilizes the Zarr storage format to enable the segmentation of extremely large images.

Installation:

pip install git+https://github.com/laprade117/interactive-unet

Usage

After installation you can run the tool using:

interactive-unet --project_folder "path/to/project_folder" --num_classes 2

This will create a project folder at the specified location then setup a user interface for a two-class segmentation task and provide a link that can be opened in any web browser. If you wish to segment more than two classes, increase the --num_classes argument to a larger value (maximum 10). A random port is used when creating the interface, however, it can be specified with --port 9090, if needed. Use interactive-unet --help for a full list of optional arguments.

The tool expects data to be stored in multi-scale Zarr files with a 'uint8' datatype. Chunk sizes of 32x32x32 or 64x64x64 have been found to be optimal for efficient navigation with the tool. See the Converting_to_zarr.ipynb notebook for some scripts for converting to the correct formats.

Keyboard Shortcuts

• Left click: Paint displayed color

• Shift + Left Click: Push displayed overlay onto annotation map

• Mouse Wheel: Adjust brush size

• C: Cycles through colors

• D: Toggle prediction overlay

• Ctrl + Z: Undo last paint stroke

• Ctrl + Y: Redo last paint stroke

• Ctrl + Left Click + Drag: Translation

• Ctrl + Mouse Wheel: Zoom in and out

• Ctrl + Right Click + Drag: Scroll through slices

• Ctrl + Middle Mouse Button + Drag: Rotate plane

Local Setup (with Conda)

1. Create a conda environment: conda create --name unet python=3.11

2. Activate the environment: conda activate unet

3. Install the tool: pip install git+https://github.com/laprade117/interactive-unet

4. Launch the tool: interactive-unet --project_folder "path/to/project_folder" --num_classes 2

DTU Thinlinc Setup (with Conda)

1. Activate an interactive GPU session: sxm2sh -X

2. Create a conda environment: conda create --name unet python=3.11

3. Activate the environment: conda activate unet

4. Install the tool: pip install git+https://github.com/laprade117/interactive-unet

5. Launch the tool: interactive-unet --project_folder "path/to/project_folder" --num_classes 2

While running on Thinlinc, you can only access the tool from the browser within the the ThinLinc Client. Access outside of the ThinLinc client is not working.