The Brain Segmentation project is focused on developing a lightweight model for the segmentation of brain images. This project involves comparing different neural network architectures, selecting the UNet model, building a training script, integrating a W&B tracker, exploring model optimizations, and conducting the final training.
-
Comparing Neural Network Architectures
- We evaluated various neural network architectures and ultimately chose the UNet model for its superior performance in segmentation tasks.
-
Building a Training Script
- We developed a robust training script to streamline the training process.
-
Adding W&B Tracker
- Integrated Weights & Biases (W&B) tracker for monitoring and visualizing the training process.
-
Exploring Model Optimizations
- Implemented and tested various optimizations to enhance model performance. For detailed steps, refer to Optimizations.
-
Final Training
- Conducted the final training phase using the optimized model and training script.
For detailed steps, refer to the following markdown files:
- NN Architectures Investigation
- Investigation Results
- Setup Jupyter
- Pipenv Environment Setup
- Train Job
- GitHub Workflow
We used the brain tumor segmentation dataset from a BRCA competition. This dataset provided the necessary data for training and evaluating our model.
The compute resources for this project were provided by the Berlin Institute of Health (BIH) during our internship.
Distributed under the MIT License. See LICENSE for more information.
- Dzmitry Hramyka
- Mathias Husted
- Evin Demir
- Nils Bender
We would like to thank the following individuals and organizations for their support and contributions:
- Berlin Institute of Health (BIH)
- Soren Lukassen (Supervisor)