This repository represents our lab's submission to the ATLAS 2.0 segmentation challenge (incorporated into ISLES at MICCAI 2022) for stroke lestion segmentation based on T1-weighted (T1w) imaging.
In our solution, we used a patch-based 3D residual U-Net to segment stroke lesions using two input channels: one being the original T1w image, and another being the same image flipped symmetrically across the left-right axis. Given the often-unilateral nature of stroke lesions, incorporation of a symmetrically flipped channel – such that the model sees a patch from roughly the same region of each hemisphere – helped improve sensitivity and overall performance significantly.
This repository was based on the ATLAS 2.0 sample Docker. The text from the original README.md is maintained below for completeness:
This repository serves as a template for your to produce a Docker container with your model.
Your model should be trained and loadable at this stage.
There are three important files: for you to modify:
requirements.txt
- Python dependencies for your model.
Python packages specified inrequirements.txt
will be installed in the container's Python environment when it is built.process.py
- Modify the section to load and call your model.
Load your model and use it to make predictions on the input.Dockerfile
- Add the files needed to run your model (model weights, code, etc.)
Once complete, you can run build.sh
to build the container, and export.sh
to package it for upload.
The original source code for the algorithm container was generated with evalutils version 0.3.1.