Semi-Automatic Brain Tumor Segment Annotation Tool using DL

We create CAD system which uses the best of both world i.e. human and computer to speeden UP the process of brain tumor segmentation.

Comment

Inspiration

Being part of brain tumor segmentation team at university, this idea has been lingering behind in my head for reasonable amount of time to create a semi automated tool for annotation purposes in medical field in which a rough work is done by the computer and only the precision work is done by human experts.

What it does

It provides a computer aided brain tumor segmentation annotation tool for radiologists to reduce the errors and help with their work.

How we built it

Four main components on how I built of the project

  • Model: Holds details regarding the model that was trained and other details. Model is written in pytorch(UNET) and used figshare dataset. This was the first step.
  • segment_server: Flask server that serves the inference model using REST api. This was the second setup and was relatively easy.

  • vectorization_server: NodeJs-Express server that converts Raster images to Vector Image svg path. These svg path are the masks that can can overlaid on the orginal tumor for annotation purpose. This was horribly difficult as no good tools available in Python. This was the third step

  • annotation_client: Dash app that provides the front-end to work with the segmentation server and allows for manual annotation after automatic annotation is done by inference model.

Challenges we ran into

By far, the biggest was conversion of raster images to svg paths with no support available in python or julia (languages I'm comfortable in). That easily eat up half my time, then I had to create a vectorization_server rest api for conversion from raster images to vector images using potrace.js to get the svg paths in a way that was acceptable to dash app and plotly graphs.

Accomplishments that we're proud of and What we learned

I never did Flask server, Express server and not even Dash server ever before. I knew about brain tumor segmentation but not enough about these topics but I think I did fairly well with my deployements.

What's next for DL supported Brain Annotation Tool

Improvements in UI,UX is biggest one and to make it more accessible. Then availabilty of other models and improvements in accuracy. There is lot of space for improvement in speed of the process.

Share this project:

Updates