MIC-DKFZ
Having similar Network??
Hi I am interested in your cardiac segmentation network, The 2D network you you shared here under the repository of MIC-DKFZ/ACDC2017 is "similar" to that you discussed in paper on give link https://arxiv.org/abs/1707.00587.
your response would be appreciable.
Hi, this is the exact code of that paper. But it is very old and we have outperformed it in the meantime. I strongly recommend you go have a look at https://github.com/MIC-DKFZ/nnUNet. There you can also find pretrained models for the ACDC and MnMs dataset Best, Fabian
Hello, Fabian, Are the networks 2D and 3D u-net integrated together (means 2D conv and 3D conv are Concatenated) or they are separated? Because it will decide the input data (2D or 3D), and training (training ensemble or one after one). I wonder the final predicted results are based on which model? Thank you. ❤️
HI, are you asking about nnU-Net or this repository? In both repositories the networks act separately. I forgot how this repo works and I would recommend not to use it. nnU-Net can automatically select what the best combination is and it will print the commands you need to run prediction. After training the 5 folds for both configurations (3d_fullres and 2d) simply run nnUNet_find_best_configuration and it will tell you which commands you need for inference
(/home/fabian/pytorch-v1.7.1_cuda-11.1_cudnn-8.1.0.77/conda_env) fabian@Fabian:~$ nnUNet_find_best_configuration -h
usage: This is intended to identify the best model based on the five fold cross-validation. Running this script requires all models to have been run already. This script will summarize the results of the five folds of all models in one json each for easy interpretability
optional arguments: -h, --help show this help message and exit -m MODELS [MODELS ...], --models MODELS [MODELS ...] -t TASK_IDS [TASK_IDS ...], --task_ids TASK_IDS [TASK_IDS ...] -tr TR nnUNetTrainer class. Default: nnUNetTrainerV2 -ctr CTR nnUNetTrainer class for cascade model. Default: nnUNetTrainerV2CascadeFullRes -pl PL plans name, Default: nnUNetPlansv2.1 -f FOLDS [FOLDS ...], --folds FOLDS [FOLDS ...] Use this if you have non-standard folds. Experienced users only. --disable_ensembling Set this flag to disable the use of ensembling. This will find the best single configuration for each task. --disable_postprocessing Set this flag if you want to disable the use of postprocessing
Please do not mess with the data. Just give nnU-Net the 3D images for training and inference. nnU-net will take care of the rest. Best, Fabian