AttentionBased-MIS
Medical Image Segmentation Based on Attention mechanism(Fine-Tune for natural image[semantic/instance] segmentation).
This project is dedicated to
- Collecting and re-implementing basic models and different attention mechanisms, transforming them modular and portable.
- Proposing novel attention mechanisms tailed for 3D data Segmentation.
Main purpose is used in 3D Medical Image Segmentation. Fine-tune for Other CV tasks need attention is easily meanwhile.
data
Base dataset and derived dataset...Coming Soon
models
This section include basic model(for segmentation or feature extraction) and different attention mechanisms.Each attention mechanism can recalibrate multi-dim feature maps across their own functional domain.
Most attention mechanisms can be modularized and integrated into any sub feature maps(e.g. each encoder in UNet or each block in ResNet)
if not special noted. so models
has basic models and attention modules. You can combine Model Name with Attention Module Name to
construct your own concrete model, for example:
--basic model ResNet --attention module CBAM
--basic model 3D UNet --attention module AG
... ...
You can Specific where the attention modules inserted in, Default is after each block/encoder/decoder.
Basic model
3D UNet: paper|reimplemented code: Coming Soon Model Name: 3D UNet
VNet: paper|reimplemented code: Coming Soon Model Name: VNet
DeepMedic: paper|reimplemented code: Coming Soon Model Name: DeepMedic
H-DenseUNet: paper|reimplemented code: Coming Soon Model Name: H-DenseUNet
VoxResNet: paper|reimplemented code: Coming Soon Model Name: VoxResNet
U-Net: paper|reimplemented code: Coming Soon Model Name: U-Net
ResNet: paper|code Model Name: ResNet
FCN: paper|reimplemented code: Coming Soon Model Name: FCN
DeepLabV3+: paper|reimplemented code: Coming Soon Model Name: DeepLabV3+
Attention module
Class Activation Map: paper|reimplemented code: Coming Soon Attention Module Name: CAM
notes: Coming soon
Spatial Transformer Net: paper|reimplemented code: Coming Soon Attention Module Name: STN
notes: Coming soon
Squeeze-and-Excitation: paper|reimplemented code: Coming Soon Attention Module Name: SE
notes: Coming soon
CBAM: paper|reimplemented code: Coming Soon Attention Module Name: CBAM
notes: Coming soon
Dual Attention: paper|reimplemented code: Coming Soon Attention Module Name: DN
notes: Coming soon
Split Attention:paper|reimplemented code: Coming Soon Attention Module Name: SpA
notes: Coming soon
Project&Excitation: paper|reimplemented code: Coming Soon Attention Module Name: PE
notes: Coming soon
Attention U-Net: paper|reimplemented code: Coming Soon Attention Module Name: AG
notes: Coming soon
Volumetric Attention: paper|reimplemented code: Coming Soon Attention Module Name: VA
notes: Coming soon
Feature Correlation Attention: paper|reimplemented code: Coming Soon Attention Module Name: FCA
notes: Coming soon
Hierarchical Attention Net: paper|reimplemented code: Coming Soon Attention Module Name: HAN
notes: Coming soon
Ours: paper: Coming Soon|Source code: Coming Soon Attention Module Name: ***
notes: Coming soon
Above models and attention modules have been experimented and still many other models waiting for test. This project will be update consistently and Welcome to advise good base model or attention modules.
options
Base options and derived options...Coming Soon
utils
Classes of Visulization,Loss,Metric,Statistic Test...Coming Soon
How to Train and Test?
**Coming Soon**