Official code for the ISBI 2024 paper: SimSAM: Zero-shot Medical Image Segmentation via Simulated Interaction, which was awarded as a top 12 oral paper.
- datasets
- monai
- numpy
- opencv-python
- tifffile
- torch
- transformers
The relevant packages can be installed with:
pip install -r requirements.txt
This repo also uses surface-distance, which should be downloaded and included in the root folder.
This paper makes use of three datasets, which need to be separately downloaded.
- Breast Ultrasound Scans can be downloaded from Kaggle.
- CVC ClinicDB is available here.
- ISIC 2016 is available here (you should download Task 1).
Downloaded content should be placed in the
data/
folder.
You can evaluate either the baseline
or the simsam
model with the eval.py
script, by selecting from one of the following datasets: ["busi", "cvc", "isic"]
. For example:
python scripts/eval.py --model_load_path facebook/sam-vit-base \
--dataset cvc \
--model_type simsam \
You can also fine-tune the SAM model on any of the datasets with:
python scripts/train.py --model_load_path facebook/sam-vit-base \
--model_save_path path/to/save/model \
--dataset cvc \
--learning_rate 1e-5 \
--num_train_epochs 10 \