M-SAM [Paper]
We introduce Mask-Enhanced SAM (M-SAM), an innovative architecture tailored for 3D tumor lesion segmentation. This method is elaborated on the paper Mask-Enhanced Segment Anything Model for Tumor Lesion Semantic Segmentation.
- We introduce a novel Mask-Enhanced SAM (M-SAM) architecture to explore the application of SAM in the medical domain, validating its effectiveness in tumor lesion segmentation.
- We propose a Mask-Enhanced Adapter (MEA) to align the positional information of the prompt with the semantic information of the input image, optimizing precise guidance for mask prediction. Based on the design of the MEA, we further implement an iterative refining scheme to refine masks, yielding improved performances.
- With updates to only about 20% of the parameters, our model outperforms state-of-the-art medical image segmentation methods on five tumor lesion segmentation benchmarks. Additionally, we validate the effectiveness of our method in domain transferring.
M-SAM consists of multiple stages of iterative refinement, which makes it possible to refine the predicted segmentation masks iteratively, thus obtaining more accurate segmentation boundaries progressively.
Install the environment:
pip install -r requirements.txtThen download SAM checkpoint, and put it at .work_dir/SAM/
- Prepare your own train dataset and refer to the samples in
M-SAM/demoto replace them according to your specific scenario. The target file structures should be like the following:
data/train ├── ct_BraTS │ ├── imagesTr │ │ ├── word_0025.nii.gz │ │ ├── ... │ ├── labelsTr │ │ ├── word_0025.nii.gz │ │ ├── ... │ ── ct_LiTS │ ├── imagesTr │ │ ├── word_0025.nii.gz │ │ ├── ... │ ├── labelsTr │ │ ├── word_0025.nii.gz │ │ ├── ... ├── ...
- Start Training!
cd ./M-SAM
python train_msam.pyThis will start the training process of the M-SAM model on your prepared data. The key options are listed below:
- task_name: task name
- checkpoint: pre-trained checkpoint
- work_dir: results folder for log and ckpt
- num_epochs: number of epoches
- batch_size: batch size for training
- num_workers: the num_workers
- lr: learning rate for training
- lr_scheduler: the choices of scheduler, default is multisteplr
💡Note: Use the --checkpoint to set the pre-trained weight path, if no checkpoint is found in the specified path, the model will undergo training from scratch.
- Evaluation
Prepare your own test dataset and refer to the samples in
M-SAM/demoto replace them according to your specific scenario. The target file structures should be like the following:
data/test ├── adrenal │ ├── ct_WORD │ │ ├── imagesTr │ │ │ ├── word_0027.nii.gz │ │ │ ├── ... │ │ ├── labelsTr │ │ │ ├── word_0027.nii.gz │ │ │ ├── ... ├── liver │ ├── ct_WORD │ │ ├── imagesTr │ │ │ ├── word_0029.nii.gz │ │ │ ├── ... │ │ ├── labelsTr │ │ │ ├── word_0029.nii.gz │ │ │ ├── ... ├── ...
And
cd ./M-SAM
python test_msam.pyThis will start the testing process of the M-SAM model on your prepared data.
@inproceedings{shi2024mask,
title={Mask-enhanced segment anything model for tumor lesion semantic segmentation},
author={Shi, Hairong and Han, Songhao and Huang, Shaofei and Liao, Yue and Li, Guanbin and Kong, Xiangxing and Zhu, Hua and Wang, Xiaomu and Liu, Si},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={403--413},
year={2024},
organization={Springer}
}
- We thank all the open-source of the following projects:
This project is released under the Apache 2.0 license.