Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
root acfa1a587b | 1 year ago | |
---|---|---|
dataset | 1 year ago | |
model | 1 year ago | |
utils | 1 year ago | |
.gitignore | 1 year ago | |
LICENSE | 1 year ago | |
readme.md | 1 year ago | |
train_mtunet_ACDC.py | 1 year ago | |
train_mtunet_Synapse.py | 1 year ago |
The paper has been accepted by ICASSP 2022. The complete code is released today.
Please note that, if you have requested our code before, that code is depreciated right now and you are encouraged to use the newest version in this repo.
1. Prepare your dataset.
Synapse dataset can be found at the repo of TransUnet.
ACDC dataset is a little complicated, since we found that previous works uses different partition but compares with each other directly. To make sure our experiment is more fair, we uses our own partition and rerun all the methods by ourselves. We have uploaded the preprocessed ACDC dataset here, or you can download the dataset by yourself at here
2. Clone the code
git clone git@github.com:Dootmaan/MT-UNet.git
cd MT-UNet
3. Start training
CUDA_VISIBLE_DEVICES=0 nohup python3 -u train_mtunet_ACDC.py >train_mtunet_ACDC.log 2>&1 &
The weights will be saved to "./checkpoint/ACDC/mtunet" while the predictions will be saved to "./predictions" by default. You can also load our weights before training with the argparser.
We have tested the code to make sure it works. However, if you still find some bugs, feel free to make a pull request or simply raise an issue.
You are also encouraged to read the update log below to know more about this repo.
By another round of training based on previous weights, our model also achieved a better performance on ACDC (91.61% DSC). We have changed the weights for ACDC to this newest version and you can check it out for yourself. However, previous versions of weights are still available on Google Drive, and you can access them via previous commits.
We have further trained our MT-UNet and it turns out to have a better result on Synapse with 79.20% DSC. We have changed the public weights of Synapse to this version and will also update the results in our paper.
It should be mentioned that we are currently conducting some statistical evaluations on our model and these results will be also made public on this site.
[Updated] Click here for our weights used on Synapse.
[Updated] Click here for our weights used on ACDC. The authors of TransUnet did not provide the split of ACDC dataset. Therefore, we conducted all the ACDC experiments based on our own dataset split.
This is the official implementation for our ICASSP2022 paper MIXED TRANSFORMER UNET FOR MEDICAL IMAGE SEGMENTATION
The entire code will be released upon paper publication.
No Description
Python
Dear OpenI User
Thank you for your continuous support to the Openl Qizhi Community AI Collaboration Platform. In order to protect your usage rights and ensure network security, we updated the Openl Qizhi Community AI Collaboration Platform Usage Agreement in January 2024. The updated agreement specifies that users are prohibited from using intranet penetration tools. After you click "Agree and continue", you can continue to use our services. Thank you for your cooperation and understanding.
For more agreement content, please refer to the《Openl Qizhi Community AI Collaboration Platform Usage Agreement》