Author: Jizong Peng Data: 18.11.2018

The code here is for the paper "DCSegNet: A Discretely-Constrained Segmentation Network for Medical Segmentation", accepted by Medical Imaging meets NIPS, NeurIPS 2018, Montreal.

Before running script, install the package by pip install -e .

For supervised training using ACDC dataset:
python train.py --dataroot=cardiac --method=fullysupervised --data_aug=True --arch=enet --loss=cross_entropy --max_epoch=1 --save_dir=results/cardiac/FS_enet_Daug --num_admm_innerloop=1 For weakly-supervised training with ACDC dataset and ADMM_size method:
python train.py --dataroot=cardiac --method=admm_size --data_aug=True --arch=enet --max_epoch=1 --save_dir=results/cardiac/size_enet_Daug_0.0 --eps=0.0 Where the eps is the size error ratio between 0 and 1.

You can set a global size constraint by using the tag: --individual_size_constraint=False and --global_upbound=20, global_lowbound==2000, as well.

For a training using GC and size together:
python train.py --dataroot=cardiac --method=admm_gc_size --data_aug=True --arch=enet --max_epoch=1 --save_dir=results/cardiac/gcsize_enet_Daug_0.0 --eps=0.0

For ADMM training with inequality constraints, using python train_in.py --dataroot=cardiac --method=admm_gc_size_in --data_aug=True --arch=enet --max_epoch=1 --save_dir=results/cardiac/gcsizeIN_enet_Daug_0.0 --eps=0.0

##Attention

Each script can cost 3 days for a whole training and the results can be different with different initializations. This ADMM way is time costing and relatively hard to tune. The hyparameters are set as a grid search of using parameterSearch.py given --name and --output_dir.

Good luck with the fine-tuning and Happy new Year.

Jizong