A tool to enable biologists to train machine learning models for their own histo-pathology images.
A short introduction video can be seen here:
Open the example projects in the client to learn how to use the software. Have a look at the Model Application and Training and Annotation Modes User Guides for a detailed textual explanation.
See the HistoFlow-Models repository for pre-trained models.
The code is released under the MIT license.
Note: The development name of the tool was luminosus-microscopy which may still appear in the code.
Windows and macOS binaries are available in the releases area at the top right. They already contain all required dependencies. Setting up or having access to a HistoFlow server is required to train or apply models (see below).
- Win / macOS / Linux (cross-compilation for Android and iOS should work, too)
- Qt 5.14
- QtCreator or qmake
You also need to create an SSL certificate (core/data/luminosus_websocket.cert and core/data/luminosus_websocket.key). Please use the standard commands you can find on the Internet to create those.
git submodule update --init --recursive- open
src/luminosus-microscopy.proin Qt Creator, configure and hit the green play button
- Python 3 and dependencies as listed in server/requirements.txt (most notably fast.ai and Flask)
- Nvidia GPU recommended for accelerated training and inference
- copy SSL certificate from client in case you want to use HTTPS
- (optionally) create and activate Conda environment
- change to the
serverdirectory - install Python dependencies (e.g. using
conda install --file requirements.txt) - run
export FLASK_APP=main.py - start the server with
python3 -m flask run --host='::' --port=55712to listen for IPv6 HTTP requests - alternatively run
python3 -m flask run --host='::' --port=55712 --cert=luminosus_websocket.cert --key=luminosus_websocket.keyfor HTTPS - enter the IP address of the server in the clients settings, it will connect to it automatically
The screenshot below shows how a CNN model is applied in the node-based GUI: The 'DAPI Image' with ~10k cells serves as the input for the model. The 'CNN Model' block represents an already trained network. The 'Area' block allows to limit the analysis area. After clicking 'Run', the 'CNN Inference' block prepares the input data (with a preview in the top part of the block), uploads it to the back end and starts the inference. The result is stored in the 'Dataset' block on the right. Under it is the output image of the network that can be used for further post-processing. A snippet of the model output is displayed as an overlay on the DAPI image at the top.
The following screenshot shows the whole user interface. Multiple images and plots can be displayed simultaneously.
Please cite this pre-print when you refer to this tool: HistoFlow Paper