This is a Tensorflow reimplemention of Dual-Stream Pyramid Registration Network
The packages and their corresponding version we used in this repository are listed in below.
- Tensorflow==1.15.4
- Keras==2.3.1
- tflearn==0.5.0
After configuring the environment, please use this command to train the model.
python train.py -g 0 --batch 1 -d datasets/brain.json -b DUAL -n 1 --round 10000 --epoch 10Use this command to obtain the testing results.
python predict.py -g 0 --batch 1 -d datasets/brain.json -c weights/Dec09-1849We use the same training and testing data as RCN, please refer to their repository to download the pre-processed data.
| Method | Dice | HD | ASSD | Jacobian Std. | Folding (%) |
|---|---|---|---|---|---|
| Original Dual-PRNet | 0.778 | - | - | - | - |
| Re-implemented Dual-PRNet | 0.831±0.008 | 3.457±0.297 | 0.811±0.046 | 0.906±0.059 | 1.6e-1±2.4e-2 |
| VoxelMorph | 0.820±0.008 | 3.648±0.284 | 0.892±0.047 | 0.247±0.057 | 5.2e-3±6.8e-3 |
| VTN | 0.825±0.008 | 3.584±0.265 | 0.925±0.047 | 0.179±0.024 | 0.0±0.0 |
| 2×10-cascade VTN | 0.831±0.009 | 3.551±0.328 | 0.810±0.046 | 0.355±0.068 | 1.2e-6±7.5e-6 |
We have tried to follow Dual-PRNet to merge 56 regions into 7, and the merged 7 regions and the corresponding label IDs of functional areas in each merged region are shown in figure below.
Some codes are modified from RCN and VoxelMorph. Thanks a lot for their great contribution.