Liver segmentation models
Table of content
Architectures
- Unet
- Unet++
- FPN
- DAF
Experiment results
We test these four segmentation models at a liver datasets. We rewrite the framework with reference to DAF. The model implementations of Unet,FPN and DAF are dependent on github while Unet++ is implemented by ourself. For fair, we train all models with 60 epochs and add evaluation for each epoch. Then save best epoch as the last result. We evaluate them with two norm: Dice and F1, all results as follows:
Model | Dice | F1 | Backbone | Batch size | Loss function | Resize | Use pretrained | Rcf refine |
---|---|---|---|---|---|---|---|---|
Unet | 0.9494 | 0.9503 | Resnet-18 | 4 | bcd,dice | (448,448) | Y | Y |
Unet++ | 0.9453 | 0.9465 | Resnet-18 | 8 | bcd,dice | (224,224) | Y | Y |
FPN | 0.9493 | 0.9491 | Resnet-18 | 4 | bcd,dice | (448,448) | Y | Y |
DAF | 0.9515 | 0.9515 | ResNext-101 | 4 | bcd,dice | (448,448) | Y | Y |