kleinzcy / sassnet Goto Github PK

View Code? Open in Web Editor NEW

155.0 5.0 28.0 66.72 MB

Shape-aware Semi-supervised 3D Semantic Segmentation for Medical Images

Python 100.00%

sassnet's Introduction

SASSnet

Code for paper: Shape-aware Semi-supervised 3D Semantic Segmentation for Medical Images(MICCAI 2020)

Our code is origin from UA-MT

You can find paper in Arxiv.

Usage

Clone the repo:

git clone https://github.com/kleinzcy/SASSnet.git 
cd SASSnet

Put the data in data/2018LA_Seg_Training Set.
Train the model

cd code
# for 16 label
python train_gan_sdfloss.py --gpu 0 --label 16 --consistency 0.01 --exp model_name
# for 8 label
python train_gan_sdfloss.py --gpu 0 --label 8 --consistency 0.015 --exp model_name

Params are the best setting in our experiment.

Test the model

python test_LA.py --model model_name --gpu 0 --iter 6000

Our best model are saved in model dir.

Citation

If you find our work is useful for you, please cite us.

sassnet's People

Contributors

Stargazers

Watchers

sassnet's Issues

data

Your work is very good.I am very eager to download the data you use in your work.But the data can no longer be downloaded from the official website.Would you please share it with me?

dataset

Hello,thank you for your excellent work! I have a question about the dataset: what is the format for training and testing images in the data folder ? I am looking forward to your reply!

high-quality SDM predictions on the labeled set

emm...how could i get the high-quality SDM predictions on the labeled set?

comparative experiment

Dear author：

Hi, Thanks for your great work in 《Shape-aware Semi-supervised 3D SemanticSegmentation for Medical Images》!

This article has benefited me a lot.

I am writing to you this time mainly to think about the code you are requesting for a comparative experiment.

DAP、 ASDNet 、TCSE

I can't find the relevant code on the Internet, so I'm very much looking forward to getting your comparative experiment code.

Looking forward to hearing from you, I wish you all the best.

Questions about the effect of SDF on 2D images

Hello, I am very sorry to disturb you. First of all, thank you very much for your excellent work, and I have a question. Your experiments were done on 3D images, and they have excellent results. I would like to ask you if this SDF map will work on 2D images, and I feel that this distance map should also work on 2D images, right? I hope you can answer me, thank you very much!

about the test

Dear author：

Hi, Thanks for your great work! I am interested in your work.

Sorry to disturb you.I have some questions. I downloaded the "2018 Atrial Segmentation Challenge COMPLETE DATASET",and then used the "la_heart_processing.py" file to process the "Training Set", and put in in the corresponding directory，and the file “train_gan_sdfloss” work on, training success. But when I test with the train model "iter_6000.pth", the result showed :average metric is [ 0.18091923 0.10030632 53.18094769 23.59822443]. If I test with your "best.pth" it can get the same result as your paper showed.

Can you give me some answers? Thanks

Best wish!

about class number

Hi, thanks for your reply, I meet a question in the code, can you help,
num_classes = 2, but during training, the input class number is n_classes=num_classes-1:
net = VNet(n_channels=1, n_classes=num_classes-1, normalization='batchnorm', has_dropout=True)
Can you tell me why do this operation?
Many thanks, waiting for your reply.

Question of your VNet result

Dear author,

I ran the baseline framework Vnet (highlights in picture below) based on the 8 label supervised protocol, and find that it is far behind the result on your table.

What I have is: 75.4014 61.832541 9.92698021 33.91709438.
I'm wondering whether you have this supervised result from the Bayesian VNet, or from your model (i.e., with the assistance of the discriminator).

kind regards,

How to generate SDMs for multi-class segmentation task?

RT.