The corresponding code for my master's thesis 'Segmentation of sparse annotated data: application to cardiac imaging'.
Initially, we built a custom pipeline using our own model for our experiments. Corresponding code for training on our custom pipeline can be run from src/main.py
. Hyperparameters and other variables can be controlled from within config.toml
in the root directory. Pass the CLI argument -d
as 'acdc' or 'mnms' depending on the desired dataset. Your data should be setup using the same pre-processing format as nnUNet.
The code for running/training nnUNet can be found in the fork at https://github.com/joshestein/nnUNet/tree/limited_data. The setup/evaluation/inference is similar to the original repo. There are some changes made to include our additional evaluations - see evaluation/evaluate_all.py
, evaluation/surface_metrics.py
and inference/predict_all.py
for some of our important changes.
Code for training our SAM models can be found within src/sam
. To train models use src/sam/sam_main.py
. Once trained, inference results can be obtained using src/sam/sam_inference.py
.